Project import generated by Copybara.

GitOrigin-RevId: da3cb82c964457b1c719c915a48e545b980af64b

.bazelrc

@@ -12,9 +12,9 @@ build --copt='-Wno-comment'
 build --copt='-Wno-return-type'
 build --copt='-Wno-unused-local-typedefs'
 build --copt='-Wno-ignored-attributes'
-# Temporarily set the incompatibility flag for Bazel 0.27.0 and above
-build --incompatible_disable_deprecated_attr_params=false
-build --incompatible_depset_is_not_iterable=false
+
+# Tensorflow needs remote repo
+build --experimental_repo_remote_exec
 
 # Sets the default Apple platform to macOS.
 build --apple_platform_type=macos
@@ -23,6 +23,10 @@ build --apple_platform_type=macos
 build --apple_generate_dsym
 
 # Android configs.
+# Note: the documentation tells us to use @androidndk//:default_crosstool, but
+# the automatic configuration transition uses //external:android/crosstool.
+# Using it here works and spares us from having two different config_settings
+# for Android.
 build:android --crosstool_top=//external:android/crosstool
 build:android --host_crosstool_top=@bazel_tools//tools/cpp:toolchain
 build:android --linkopt=-landroid

Dockerfile

@@ -48,7 +48,7 @@ RUN pip install future
 RUN pip3 install six
 
 # Install bazel
-ARG BAZEL_VERSION=1.1.0
+ARG BAZEL_VERSION=2.0.0
 RUN mkdir /bazel && \
     wget --no-check-certificate -O /bazel/installer.sh "https://github.com/bazelbuild/bazel/releases/download/${BAZEL_VERSION}/b\
 azel-${BAZEL_VERSION}-installer-linux-x86_64.sh" && \

README.md

@@ -15,6 +15,7 @@
 * [Hair Segmentation](mediapipe/docs/hair_segmentation_mobile_gpu.md) [[Web Demo]](https://viz.mediapipe.dev/runner/demos/hair_segmentation/hair_segmentation.html)
 * [Object Detection](mediapipe/docs/object_detection_mobile_gpu.md)
 * [Object Detection and Tracking](mediapipe/docs/object_tracking_mobile_gpu.md)
+* [Objectron: 3D Object Detection and Tracking](mediapipe/docs/objectron_mobile_gpu.md)
 * [AutoFlip](mediapipe/docs/autoflip.md)
 
 ![face_detection](mediapipe/docs/images/mobile/face_detection_android_gpu_small.gif)
@@ -43,6 +44,8 @@ A web-based visualizer is hosted on [viz.mediapipe.dev](https://viz.mediapipe.de
 *  [YouTube Channel](https://www.youtube.com/channel/UCObqmpuSMx-usADtL_qdMAw)
 
 ## Publications
+* [MediaPipe Objectron: Real-time 3D Object Detection on Mobile Devices](https://mediapipe.page.link/objectron-aiblog)
+* [AutoFlip: An Open Source Framework for Intelligent Video Reframing](https://mediapipe.page.link/autoflip)
 * [Google Developer Blog: MediaPipe on the Web](https://mediapipe.page.link/webdevblog)
 * [Google Developer Blog: Object Detection and Tracking using MediaPipe](https://mediapipe.page.link/objecttrackingblog)
 * [On-Device, Real-Time Hand Tracking with MediaPipe](https://ai.googleblog.com/2019/08/on-device-real-time-hand-tracking-with.html)
@@ -63,7 +66,7 @@ A web-based visualizer is hosted on [viz.mediapipe.dev](https://viz.mediapipe.de
 *  [Discuss](https://groups.google.com/forum/#!forum/mediapipe) - General community discussion around MediaPipe
 
 ## Alpha Disclaimer
-MediaPipe is currently in alpha for v0.6. We are still making breaking API changes and expect to get to stable API by v1.0.
+MediaPipe is currently in alpha for v0.7. We are still making breaking API changes and expect to get to stable API by v1.0.
 
 ## Contributing
 We welcome contributions. Please follow these [guidelines](./CONTRIBUTING.md).

WORKSPACE

@@ -2,23 +2,22 @@ workspace(name = "mediapipe")
 
 load("@bazel_tools//tools/build_defs/repo:http.bzl", "http_archive")
 
-skylib_version = "0.8.0"
+skylib_version = "0.9.0"
 http_archive(
     name = "bazel_skylib",
     type = "tar.gz",
-    url = "https://github.com/bazelbuild/bazel-skylib/releases/download/{}/bazel-skylib.{}.tar.gz".format (skylib_version, skylib_version),
-    sha256 = "2ef429f5d7ce7111263289644d233707dba35e39696377ebab8b0bc701f7818e",
+    url = "https://github.com/bazelbuild/bazel-skylib/releases/download/{}/bazel_skylib-{}.tar.gz".format (skylib_version, skylib_version),
+    sha256 = "1dde365491125a3db70731e25658dfdd3bc5dbdfd11b840b3e987ecf043c7ca0",
 )
 load("@bazel_skylib//lib:versions.bzl", "versions")
-versions.check(minimum_bazel_version = "0.24.1",
-               maximum_bazel_version = "1.2.1")
+versions.check(minimum_bazel_version = "2.0.0")
 
 
-# ABSL cpp library lts_2019_08_08.
+# ABSL cpp library lts_2020_02_25
 http_archive(
     name = "com_google_absl",
     urls = [
-        "https://github.com/abseil/abseil-cpp/archive/20190808.tar.gz",
+        "https://github.com/abseil/abseil-cpp/archive/20200225.tar.gz",
     ],
     # Remove after https://github.com/abseil/abseil-cpp/issues/326 is solved.
     patches = [
@@ -27,8 +26,8 @@ http_archive(
     patch_args = [
         "-p1",
     ],
-    strip_prefix = "abseil-cpp-20190808",
-    sha256 = "8100085dada279bf3ee00cd064d43b5f55e5d913be0dfe2906f06f8f28d5b37e"
+    strip_prefix = "abseil-cpp-20200225",
+    sha256 = "728a813291bdec2aa46eab8356ace9f75ac2ed9dfe2df5ab603c4e6c09f1c353"
 )
 
 http_archive(
@@ -93,11 +92,13 @@ http_archive(
     build_file = "@//third_party:libyuv.BUILD",
 )
 
+# Note: protobuf-javalite is no longer released as a separate download, it's included in the main Java download.
+# ...but the Java download is currently broken, so we use the "source" download.
 http_archive(
     name = "com_google_protobuf_javalite",
-    sha256 = "79d102c61e2a479a0b7e5fc167bcfaa4832a0c6aad4a75fa7da0480564931bcc",
-    strip_prefix = "protobuf-384989534b2246d413dbcd750744faab2607b516",
-    urls = ["https://github.com/google/protobuf/archive/384989534b2246d413dbcd750744faab2607b516.zip"],
+    sha256 = "a79d19dcdf9139fa4b81206e318e33d245c4c9da1ffed21c87288ed4380426f9",
+    strip_prefix = "protobuf-3.11.4",
+    urls = ["https://github.com/protocolbuffers/protobuf/archive/v3.11.4.tar.gz"],
 )
 
 http_archive(
@@ -117,18 +118,18 @@ http_archive(
     ],
 )
 
-# 2019-11-21
-_TENSORFLOW_GIT_COMMIT = "f482488b481a799ca07e7e2d153cf47b8e91a60c"
-_TENSORFLOW_SHA256= "8d9118c2ce186c7e1403f04b96982fe72c184060c7f7a93e30a28dca358694f0"
+# 2020-04-01
+_TENSORFLOW_GIT_COMMIT = "805e47cea96c7e8c6fccf494d40a2392dc99fdd8"
+_TENSORFLOW_SHA256= "9ee3ae604c2e1345ac60345becee6d659364721513f9cb8652eb2e7138320ca5"
 http_archive(
     name = "org_tensorflow",
     urls = [
       "https://mirror.bazel.build/github.com/tensorflow/tensorflow/archive/%s.tar.gz" % _TENSORFLOW_GIT_COMMIT,
       "https://github.com/tensorflow/tensorflow/archive/%s.tar.gz" % _TENSORFLOW_GIT_COMMIT,
     ],
-    # Patch https://github.com/tensorflow/tensorflow/commit/e3a7bdbebb99352351a19e2e403136166aa52934
     patches = [
-        "@//third_party:org_tensorflow_e3a7bdbebb99352351a19e2e403136166aa52934.diff"
+        "@//third_party:org_tensorflow_compatibility_fixes.diff",
+        "@//third_party:org_tensorflow_protobuf_updates.diff",
     ],
     patch_args = [
         "-p1",
@@ -153,10 +154,6 @@ http_archive(
     sha256 = "5ba6d0db4e784621fda44a50c58bb23b0892684692f0c623e2063f9c19f192f1"
 )
 
-# Please run
-# $ sudo apt-get install libopencv-core-dev libopencv-highgui-dev \
-#                        libopencv-calib3d-dev libopencv-features2d-dev \
-#                        libopencv-imgproc-dev libopencv-video-dev
 new_local_repository(
     name = "linux_opencv",
     build_file = "@//third_party:opencv_linux.BUILD",
@@ -169,7 +166,6 @@ new_local_repository(
     path = "/usr"
 )
 
-# Please run $ brew install opencv@3
 new_local_repository(
     name = "macos_opencv",
     build_file = "@//third_party:opencv_macos.BUILD",
@@ -202,79 +198,6 @@ http_archive(
     url = "https://github.com/opencv/opencv/releases/download/3.2.0/opencv-3.2.0-ios-framework.zip",
 )
 
-RULES_JVM_EXTERNAL_TAG = "2.2"
-RULES_JVM_EXTERNAL_SHA = "f1203ce04e232ab6fdd81897cf0ff76f2c04c0741424d192f28e65ae752ce2d6"
-
-http_archive(
-    name = "rules_jvm_external",
-    strip_prefix = "rules_jvm_external-%s" % RULES_JVM_EXTERNAL_TAG,
-    sha256 = RULES_JVM_EXTERNAL_SHA,
-    url = "https://github.com/bazelbuild/rules_jvm_external/archive/%s.zip" % RULES_JVM_EXTERNAL_TAG,
-)
-
-load("@rules_jvm_external//:defs.bzl", "maven_install")
-
-maven_install(
-    artifacts = [
-        "androidx.annotation:annotation:aar:1.1.0",
-        "androidx.appcompat:appcompat:aar:1.1.0-rc01",
-        "androidx.camera:camera-core:aar:1.0.0-alpha06",
-        "androidx.camera:camera-camera2:aar:1.0.0-alpha06",
-        "androidx.constraintlayout:constraintlayout:aar:1.1.3",
-        "androidx.core:core:aar:1.1.0-rc03",
-        "androidx.legacy:legacy-support-v4:aar:1.0.0",
-        "androidx.recyclerview:recyclerview:aar:1.1.0-beta02",
-        "com.google.android.material:material:aar:1.0.0-rc01",
-    ],
-    repositories = [
-        "https://dl.google.com/dl/android/maven2",
-        "https://repo1.maven.org/maven2",
-    ],
-)
-
-maven_server(
-    name = "google_server",
-    url = "https://dl.google.com/dl/android/maven2",
-)
-
-maven_jar(
-    name = "androidx_lifecycle",
-    artifact = "androidx.lifecycle:lifecycle-common:2.0.0",
-    sha1 = "e070ffae07452331bc5684734fce6831d531785c",
-    server = "google_server",
-)
-
-maven_jar(
-    name = "androidx_concurrent_futures",
-    artifact = "androidx.concurrent:concurrent-futures:1.0.0-alpha03",
-    sha1 = "b528df95c7e2fefa2210c0c742bf3e491c1818ae",
-    server = "google_server",
-)
-
-maven_jar(
-    name = "com_google_guava_android",
-    artifact = "com.google.guava:guava:27.0.1-android",
-    sha1 = "b7e1c37f66ef193796ccd7ea6e80c2b05426182d",
-)
-
-maven_jar(
-    name = "com_google_common_flogger",
-    artifact = "com.google.flogger:flogger:0.3.1",
-    sha1 = "585030fe1ec709760cbef997a459729fb965df0e",
-)
-
-maven_jar(
-    name = "com_google_common_flogger_system_backend",
-    artifact = "com.google.flogger:flogger-system-backend:0.3.1",
-    sha1 = "287b569d76abcd82f9de87fe41829fbc7ebd8ac9",
-)
-
-maven_jar(
-    name = "com_google_code_findbugs",
-    artifact = "com.google.code.findbugs:jsr305:3.0.2",
-    sha1 = "25ea2e8b0c338a877313bd4672d3fe056ea78f0d",
-)
-
 # You may run setup_android.sh to install Android SDK and NDK.
 android_ndk_repository(
     name = "androidndk",
@@ -288,9 +211,15 @@ android_sdk_repository(
 
 http_archive(
     name = "build_bazel_rules_apple",
-    sha256 = "bdc8e66e70b8a75da23b79f1f8c6207356df07d041d96d2189add7ee0780cf4e",
-    strip_prefix = "rules_apple-b869b0d3868d78a1d4ffd866ccb304fb68aa12c3",
-    url = "https://github.com/bazelbuild/rules_apple/archive/b869b0d3868d78a1d4ffd866ccb304fb68aa12c3.tar.gz",
+    sha256 = "7a7afdd4869bb201c9352eed2daf37294d42b093579b70423490c1b4d4f6ce42",
+    url = "https://github.com/bazelbuild/rules_apple/releases/download/0.19.0/rules_apple.0.19.0.tar.gz",
+    patches = [
+        # Bypass checking ios unit test runner when building MP ios applications.
+        "@//third_party:build_bazel_rules_apple_bypass_test_runner_check.diff"
+    ],
+    patch_args = [
+        "-p1",
+    ],
 )
 
 load(
@@ -324,3 +253,49 @@ http_archive(
     build_file = "@//third_party:google_toolbox_for_mac.BUILD",
 )
 
+# Maven dependencies.
+
+RULES_JVM_EXTERNAL_TAG = "3.2"
+RULES_JVM_EXTERNAL_SHA = "82262ff4223c5fda6fb7ff8bd63db8131b51b413d26eb49e3131037e79e324af"
+
+http_archive(
+    name = "rules_jvm_external",
+    strip_prefix = "rules_jvm_external-%s" % RULES_JVM_EXTERNAL_TAG,
+    sha256 = RULES_JVM_EXTERNAL_SHA,
+    url = "https://github.com/bazelbuild/rules_jvm_external/archive/%s.zip" % RULES_JVM_EXTERNAL_TAG,
+)
+
+load("@rules_jvm_external//:defs.bzl", "maven_install")
+
+# Important: there can only be one maven_install rule. Add new maven deps here.
+maven_install(
+    name = "maven",
+    artifacts = [
+        "junit:junit:4.12",
+        "androidx.test.espresso:espresso-core:3.1.1",
+        "org.hamcrest:hamcrest-library:1.3",
+        "androidx.concurrent:concurrent-futures:1.0.0-alpha03",
+        "androidx.lifecycle:lifecycle-common:2.2.0",
+        "androidx.annotation:annotation:aar:1.1.0",
+        "androidx.appcompat:appcompat:aar:1.1.0-rc01",
+        "androidx.camera:camera-core:aar:1.0.0-alpha06",
+        "androidx.camera:camera-camera2:aar:1.0.0-alpha06",
+        "androidx.constraintlayout:constraintlayout:aar:1.1.3",
+        "androidx.core:core:aar:1.1.0-rc03",
+        "androidx.legacy:legacy-support-v4:aar:1.0.0",
+        "androidx.recyclerview:recyclerview:aar:1.1.0-beta02",
+        "com.google.android.material:material:aar:1.0.0-rc01",
+        "com.google.code.findbugs:jsr305:3.0.2",
+        "com.google.flogger:flogger-system-backend:0.3.1",
+        "com.google.flogger:flogger:0.3.1",
+        "com.google.guava:guava:27.0.1-android",
+    ],
+    repositories = [
+        "https://jcenter.bintray.com",
+        "https://maven.google.com",
+        "https://dl.google.com/dl/android/maven2",
+        "https://repo1.maven.org/maven2",
+    ],
+    fetch_sources = True,
+    version_conflict_policy = "pinned",
+)

mediapipe/BUILD

@@ -14,6 +14,9 @@
 
 licenses(["notice"])  # Apache 2.0
 
+# Note: yes, these need to use "//external:android/crosstool", not
+# @androidndk//:default_crosstool.
+
 config_setting(
     name = "android",
     values = {"crosstool_top": "//external:android/crosstool"},

mediapipe/calculators/core/BUILD

@@ -228,7 +228,9 @@ cc_library(
         "//mediapipe/framework:calculator_framework",
         "//mediapipe/framework:collection_item_id",
         "//mediapipe/framework:packet",
+        "//mediapipe/framework/formats:detection_cc_proto",
         "//mediapipe/framework/formats:landmark_cc_proto",
+        "//mediapipe/framework/formats:matrix",
         "//mediapipe/framework/formats:rect_cc_proto",
         "//mediapipe/framework/port:integral_types",
         "//mediapipe/framework/port:ret_check",
@@ -249,12 +251,14 @@ cc_library(
         "//mediapipe/framework:calculator_framework",
         "//mediapipe/framework:collection_item_id",
         "//mediapipe/framework:packet",
+        "//mediapipe/framework/formats:classification_cc_proto",
         "//mediapipe/framework/formats:landmark_cc_proto",
         "//mediapipe/framework/formats:rect_cc_proto",
         "//mediapipe/framework/port:integral_types",
         "//mediapipe/framework/port:ret_check",
         "//mediapipe/framework/port:status",
         "//mediapipe/util:render_data_cc_proto",
+        "@org_tensorflow//tensorflow/lite:framework",
     ],
     alwayslink = 1,
 )
@@ -265,10 +269,11 @@ cc_test(
     deps = [
         ":begin_loop_calculator",
         ":end_loop_calculator",
-        "//mediapipe/calculators/core:packet_cloner_calculator",
+        ":gate_calculator",
         "//mediapipe/framework:calculator_context",
         "//mediapipe/framework:calculator_contract",
         "//mediapipe/framework:calculator_framework",
+        "//mediapipe/framework:packet",
         "//mediapipe/framework/port:gtest_main",
         "//mediapipe/framework/port:integral_types",
         "//mediapipe/framework/port:parse_text_proto",
@@ -334,6 +339,7 @@ cc_library(
     deps = [
         ":clip_vector_size_calculator_cc_proto",
         "//mediapipe/framework:calculator_framework",
+        "//mediapipe/framework/formats:detection_cc_proto",
         "//mediapipe/framework/formats:rect_cc_proto",
         "//mediapipe/framework/port:ret_check",
         "//mediapipe/framework/port:status",
@@ -610,6 +616,22 @@ cc_library(
     alwayslink = 1,
 )
 
+cc_test(
+    name = "side_packet_to_stream_calculator_test",
+    srcs = ["side_packet_to_stream_calculator_test.cc"],
+    deps = [
+        ":side_packet_to_stream_calculator",
+        "//mediapipe/framework:calculator_framework",
+        "//mediapipe/framework/port:gtest_main",
+        "//mediapipe/framework/port:integral_types",
+        "//mediapipe/framework/port:parse_text_proto",
+        "//mediapipe/framework/port:status",
+        "//mediapipe/framework/tool:options_util",
+        "@com_google_absl//absl/memory",
+        "@com_google_absl//absl/strings",
+    ],
+)
+
 cc_test(
     name = "immediate_mux_calculator_test",
     srcs = ["immediate_mux_calculator_test.cc"],
@@ -677,15 +699,17 @@ cc_test(
     name = "previous_loopback_calculator_test",
     srcs = ["previous_loopback_calculator_test.cc"],
     deps = [
+        ":gate_calculator",
+        ":make_pair_calculator",
+        ":pass_through_calculator",
         ":previous_loopback_calculator",
-        "//mediapipe/calculators/core:make_pair_calculator",
-        "//mediapipe/calculators/core:pass_through_calculator",
         "//mediapipe/framework:calculator_framework",
         "//mediapipe/framework:calculator_runner",
         "//mediapipe/framework:timestamp",
         "//mediapipe/framework/port:gtest_main",
         "//mediapipe/framework/port:integral_types",
         "//mediapipe/framework/port:parse_text_proto",
+        "//mediapipe/framework/port:status",
         "//mediapipe/framework/stream_handler:immediate_input_stream_handler",
         "//mediapipe/framework/tool:sink",
         "@com_google_absl//absl/time",
@@ -753,9 +777,11 @@ cc_library(
     visibility = ["//visibility:public"],
     deps = [
         ":split_vector_calculator_cc_proto",
+        "//mediapipe/framework/formats:detection_cc_proto",
         "//mediapipe/framework:calculator_framework",
         "//mediapipe/framework/formats:landmark_cc_proto",
         "//mediapipe/framework/formats:rect_cc_proto",
+        "//mediapipe/framework/formats:matrix",
         "//mediapipe/framework/port:ret_check",
         "//mediapipe/framework/port:status",
         "//mediapipe/util:resource_util",

mediapipe/calculators/core/begin_end_loop_calculator_graph_test.cc

@@ -20,6 +20,8 @@
 #include "mediapipe/calculators/core/end_loop_calculator.h"
 #include "mediapipe/framework/calculator_contract.h"
 #include "mediapipe/framework/calculator_framework.h"
+#include "mediapipe/framework/packet.h"
+#include "mediapipe/framework/port/gmock.h"
 #include "mediapipe/framework/port/gtest.h"
 #include "mediapipe/framework/port/integral_types.h"
 #include "mediapipe/framework/port/parse_text_proto.h"
@@ -28,6 +30,13 @@
 namespace mediapipe {
 namespace {
 
+MATCHER_P2(PacketOfIntsEq, timestamp, value, "") {
+  Timestamp actual_timestamp = arg.Timestamp();
+  const auto& actual_value = arg.template Get<std::vector<int>>();
+  return testing::Value(actual_timestamp, testing::Eq(timestamp)) &&
+         testing::Value(actual_value, testing::ElementsAreArray(value));
+}
+
 typedef BeginLoopCalculator<std::vector<int>> BeginLoopIntegerCalculator;
 REGISTER_CALCULATOR(BeginLoopIntegerCalculator);
 
@@ -59,8 +68,8 @@ REGISTER_CALCULATOR(EndLoopIntegersCalculator);
 
 class BeginEndLoopCalculatorGraphTest : public ::testing::Test {
  protected:
-  BeginEndLoopCalculatorGraphTest() {
-    graph_config_ = ParseTextProtoOrDie<CalculatorGraphConfig>(
+  void SetUp() override {
+    auto graph_config = ParseTextProtoOrDie<CalculatorGraphConfig>(
         R"(
           num_threads: 4
           input_stream: "ints"
@@ -82,94 +91,222 @@ class BeginEndLoopCalculatorGraphTest : public ::testing::Test {
             output_stream: "ITERABLE:ints_plus_one"
           }
         )");
-    tool::AddVectorSink("ints_plus_one", &graph_config_, &output_packets_);
+    tool::AddVectorSink("ints_plus_one", &graph_config, &output_packets_);
+    MP_ASSERT_OK(graph_.Initialize(graph_config));
+    MP_ASSERT_OK(graph_.StartRun({}));
   }
 
-  CalculatorGraphConfig graph_config_;
+  void SendPacketOfInts(Timestamp timestamp, std::vector<int> ints) {
+    MP_ASSERT_OK(graph_.AddPacketToInputStream(
+        "ints", MakePacket<std::vector<int>>(std::move(ints)).At(timestamp)));
+  }
+
+  CalculatorGraph graph_;
   std::vector<Packet> output_packets_;
 };
 
+TEST_F(BeginEndLoopCalculatorGraphTest, InputStreamForIterableIsEmpty) {
+  MP_ASSERT_OK(graph_.WaitUntilIdle());
+
+  // EndLoopCalc will forward the timestamp bound because there are no packets
+  // to process.
+  ASSERT_EQ(0, output_packets_.size());
+
+  MP_ASSERT_OK(graph_.CloseAllPacketSources());
+  MP_ASSERT_OK(graph_.WaitUntilDone());
+}
+
 TEST_F(BeginEndLoopCalculatorGraphTest, SingleEmptyVector) {
-  CalculatorGraph graph;
-  MP_EXPECT_OK(graph.Initialize(graph_config_));
-  MP_EXPECT_OK(graph.StartRun({}));
-  auto input_vector = absl::make_unique<std::vector<int>>();
-  Timestamp input_timestamp = Timestamp(0);
-  MP_ASSERT_OK(graph.AddPacketToInputStream(
-      "ints", Adopt(input_vector.release()).At(input_timestamp)));
-  MP_ASSERT_OK(graph.WaitUntilIdle());
+  SendPacketOfInts(Timestamp(0), {});
+  MP_ASSERT_OK(graph_.WaitUntilIdle());
 
   // EndLoopCalc will forward the timestamp bound because there are no elements
   // in collection to output.
-  ASSERT_EQ(0, output_packets_.size());
+  EXPECT_TRUE(output_packets_.empty());
 
-  MP_ASSERT_OK(graph.CloseAllPacketSources());
-  MP_ASSERT_OK(graph.WaitUntilDone());
+  MP_ASSERT_OK(graph_.CloseAllPacketSources());
+  MP_ASSERT_OK(graph_.WaitUntilDone());
 }
 
 TEST_F(BeginEndLoopCalculatorGraphTest, SingleNonEmptyVector) {
-  CalculatorGraph graph;
-  MP_EXPECT_OK(graph.Initialize(graph_config_));
-  MP_EXPECT_OK(graph.StartRun({}));
-  auto input_vector = absl::make_unique<std::vector<int>>();
-  input_vector->emplace_back(0);
-  input_vector->emplace_back(1);
-  input_vector->emplace_back(2);
   Timestamp input_timestamp = Timestamp(0);
-  MP_ASSERT_OK(graph.AddPacketToInputStream(
-      "ints", Adopt(input_vector.release()).At(input_timestamp)));
-  MP_ASSERT_OK(graph.WaitUntilIdle());
+  SendPacketOfInts(input_timestamp, {0, 1, 2});
+  MP_ASSERT_OK(graph_.WaitUntilIdle());
 
-  ASSERT_EQ(1, output_packets_.size());
-  EXPECT_EQ(input_timestamp, output_packets_[0].Timestamp());
-  std::vector<int> expected_output_vector = {1, 2, 3};
-  EXPECT_EQ(expected_output_vector, output_packets_[0].Get<std::vector<int>>());
+  EXPECT_THAT(output_packets_,
+              testing::ElementsAre(
+                  PacketOfIntsEq(input_timestamp, std::vector<int>{1, 2, 3})));
 
-  MP_ASSERT_OK(graph.CloseAllPacketSources());
-  MP_ASSERT_OK(graph.WaitUntilDone());
+  MP_ASSERT_OK(graph_.CloseAllPacketSources());
+  MP_ASSERT_OK(graph_.WaitUntilDone());
 }
 
 TEST_F(BeginEndLoopCalculatorGraphTest, MultipleVectors) {
-  CalculatorGraph graph;
-  MP_EXPECT_OK(graph.Initialize(graph_config_));
-  MP_EXPECT_OK(graph.StartRun({}));
-
-  auto input_vector0 = absl::make_unique<std::vector<int>>();
-  input_vector0->emplace_back(0);
-  input_vector0->emplace_back(1);
   Timestamp input_timestamp0 = Timestamp(0);
-  MP_ASSERT_OK(graph.AddPacketToInputStream(
-      "ints", Adopt(input_vector0.release()).At(input_timestamp0)));
+  SendPacketOfInts(input_timestamp0, {0, 1});
 
-  auto input_vector1 = absl::make_unique<std::vector<int>>();
   Timestamp input_timestamp1 = Timestamp(1);
-  MP_ASSERT_OK(graph.AddPacketToInputStream(
-      "ints", Adopt(input_vector1.release()).At(input_timestamp1)));
+  SendPacketOfInts(input_timestamp1, {});
 
-  auto input_vector2 = absl::make_unique<std::vector<int>>();
-  input_vector2->emplace_back(2);
-  input_vector2->emplace_back(3);
   Timestamp input_timestamp2 = Timestamp(2);
-  MP_ASSERT_OK(graph.AddPacketToInputStream(
-      "ints", Adopt(input_vector2.release()).At(input_timestamp2)));
+  SendPacketOfInts(input_timestamp2, {2, 3});
 
-  MP_ASSERT_OK(graph.CloseAllPacketSources());
-  MP_ASSERT_OK(graph.WaitUntilDone());
-
-  ASSERT_EQ(2, output_packets_.size());
-
-  EXPECT_EQ(input_timestamp0, output_packets_[0].Timestamp());
-  std::vector<int> expected_output_vector0 = {1, 2};
-  EXPECT_EQ(expected_output_vector0,
-            output_packets_[0].Get<std::vector<int>>());
+  MP_ASSERT_OK(graph_.CloseAllPacketSources());
+  MP_ASSERT_OK(graph_.WaitUntilDone());
 
   // At input_timestamp1, EndLoopCalc will forward timestamp bound as there are
   // no elements in vector to process.
+  EXPECT_THAT(output_packets_,
+              testing::ElementsAre(
+                  PacketOfIntsEq(input_timestamp0, std::vector<int>{1, 2}),
+                  PacketOfIntsEq(input_timestamp2, std::vector<int>{3, 4})));
+}
 
-  EXPECT_EQ(input_timestamp2, output_packets_[1].Timestamp());
-  std::vector<int> expected_output_vector2 = {3, 4};
-  EXPECT_EQ(expected_output_vector2,
-            output_packets_[1].Get<std::vector<int>>());
+// Passes non empty vector through or outputs empty vector in case of timestamp
+// bound update.
+class PassThroughOrEmptyVectorCalculator : public CalculatorBase {
+ public:
+  static ::mediapipe::Status GetContract(CalculatorContract* cc) {
+    cc->SetProcessTimestampBounds(true);
+    cc->Inputs().Index(0).Set<std::vector<int>>();
+    cc->Outputs().Index(0).Set<std::vector<int>>();
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Open(CalculatorContext* cc) override {
+    cc->SetOffset(TimestampDiff(0));
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Process(CalculatorContext* cc) override {
+    if (!cc->Inputs().Index(0).IsEmpty()) {
+      cc->Outputs().Index(0).AddPacket(cc->Inputs().Index(0).Value());
+    } else {
+      cc->Outputs().Index(0).AddPacket(
+          MakePacket<std::vector<int>>(std::vector<int>())
+              .At(cc->InputTimestamp()));
+    }
+    return ::mediapipe::OkStatus();
+  }
+};
+
+REGISTER_CALCULATOR(PassThroughOrEmptyVectorCalculator);
+
+class BeginEndLoopCalculatorGraphProcessingEmptyPacketsTest
+    : public ::testing::Test {
+ protected:
+  void SetUp() override {
+    auto graph_config = ParseTextProtoOrDie<CalculatorGraphConfig>(
+        R"(
+          num_threads: 4
+          input_stream: "ints"
+          input_stream: "force_ints_to_be_timestamp_bound_update"
+          node {
+            calculator: "GateCalculator"
+            input_stream: "ints"
+            input_stream: "DISALLOW:force_ints_to_be_timestamp_bound_update"
+            output_stream: "ints_passed_through"
+          }
+          node {
+            calculator: "BeginLoopIntegerCalculator"
+            input_stream: "ITERABLE:ints_passed_through"
+            output_stream: "ITEM:int"
+            output_stream: "BATCH_END:timestamp"
+          }
+          node {
+            calculator: "IncrementCalculator"
+            input_stream: "int"
+            output_stream: "int_plus_one"
+          }
+          node {
+            calculator: "EndLoopIntegersCalculator"
+            input_stream: "ITEM:int_plus_one"
+            input_stream: "BATCH_END:timestamp"
+            output_stream: "ITERABLE:ints_plus_one"
+          }
+          node {
+            calculator: "PassThroughOrEmptyVectorCalculator"
+            input_stream: "ints_plus_one"
+            output_stream: "ints_plus_one_passed_through"
+          }
+        )");
+    tool::AddVectorSink("ints_plus_one_passed_through", &graph_config,
+                        &output_packets_);
+    MP_ASSERT_OK(graph_.Initialize(graph_config));
+    MP_ASSERT_OK(graph_.StartRun({}));
+  }
+
+  void SendPacketOfIntsOrBound(Timestamp timestamp, std::vector<int> ints) {
+    // All "ints" packets which are empty are forced to be just timestamp
+    // bound updates for begin loop calculator.
+    bool force_ints_to_be_timestamp_bound_update = ints.empty();
+    MP_ASSERT_OK(graph_.AddPacketToInputStream(
+        "force_ints_to_be_timestamp_bound_update",
+        MakePacket<bool>(force_ints_to_be_timestamp_bound_update)
+            .At(timestamp)));
+    MP_ASSERT_OK(graph_.AddPacketToInputStream(
+        "ints", MakePacket<std::vector<int>>(std::move(ints)).At(timestamp)));
+  }
+
+  CalculatorGraph graph_;
+  std::vector<Packet> output_packets_;
+};
+
+TEST_F(BeginEndLoopCalculatorGraphProcessingEmptyPacketsTest,
+       SingleEmptyVector) {
+  SendPacketOfIntsOrBound(Timestamp(0), {});
+  MP_ASSERT_OK(graph_.WaitUntilIdle());
+
+  EXPECT_THAT(output_packets_, testing::ElementsAre(PacketOfIntsEq(
+                                   Timestamp(0), std::vector<int>{})));
+
+  MP_ASSERT_OK(graph_.CloseAllPacketSources());
+  MP_ASSERT_OK(graph_.WaitUntilDone());
+}
+
+TEST_F(BeginEndLoopCalculatorGraphProcessingEmptyPacketsTest,
+       SingleNonEmptyVector) {
+  SendPacketOfIntsOrBound(Timestamp(0), {0, 1, 2});
+  MP_ASSERT_OK(graph_.WaitUntilIdle());
+
+  EXPECT_THAT(output_packets_, testing::ElementsAre(PacketOfIntsEq(
+                                   Timestamp(0), std::vector<int>{1, 2, 3})));
+
+  MP_ASSERT_OK(graph_.CloseAllPacketSources());
+  MP_ASSERT_OK(graph_.WaitUntilDone());
+}
+
+TEST_F(BeginEndLoopCalculatorGraphProcessingEmptyPacketsTest, MultipleVectors) {
+  SendPacketOfIntsOrBound(Timestamp(0), {});
+  // Waiting until idle to guarantee all timestamp bound updates are processed
+  // individually. (Timestamp bounds updates occur in the provide config only
+  // if input is an empty vector.)
+  MP_ASSERT_OK(graph_.WaitUntilIdle());
+
+  SendPacketOfIntsOrBound(Timestamp(1), {0, 1});
+  SendPacketOfIntsOrBound(Timestamp(2), {});
+  // Waiting until idle to guarantee all timestamp bound updates are processed
+  // individually. (Timestamp bounds updates occur in the provide config only
+  // if input is an empty vector.)
+  MP_ASSERT_OK(graph_.WaitUntilIdle());
+
+  SendPacketOfIntsOrBound(Timestamp(3), {2, 3});
+  SendPacketOfIntsOrBound(Timestamp(4), {});
+  // Waiting until idle to guarantee all timestamp bound updates are processed
+  // individually. (Timestamp bounds updates occur in the provide config only
+  // if input is an empty vector.)
+  MP_ASSERT_OK(graph_.WaitUntilIdle());
+
+  MP_ASSERT_OK(graph_.CloseAllPacketSources());
+  MP_ASSERT_OK(graph_.WaitUntilDone());
+
+  EXPECT_THAT(
+      output_packets_,
+      testing::ElementsAre(PacketOfIntsEq(Timestamp(0), std::vector<int>{}),
+                           PacketOfIntsEq(Timestamp(1), std::vector<int>{1, 2}),
+                           PacketOfIntsEq(Timestamp(2), std::vector<int>{}),
+                           PacketOfIntsEq(Timestamp(3), std::vector<int>{3, 4}),
+                           PacketOfIntsEq(Timestamp(4), std::vector<int>{})));
 }
 
 class MultiplierCalculator : public CalculatorBase {
@@ -199,8 +336,8 @@ REGISTER_CALCULATOR(MultiplierCalculator);
 
 class BeginEndLoopCalculatorGraphWithClonedInputsTest : public ::testing::Test {
  protected:
-  BeginEndLoopCalculatorGraphWithClonedInputsTest() {
-    graph_config_ = ParseTextProtoOrDie<CalculatorGraphConfig>(
+  void SetUp() override {
+    auto graph_config = ParseTextProtoOrDie<CalculatorGraphConfig>(
         R"(
           num_threads: 4
           input_stream: "ints"
@@ -226,109 +363,85 @@ class BeginEndLoopCalculatorGraphWithClonedInputsTest : public ::testing::Test {
             output_stream: "ITERABLE:multiplied_ints"
           }
         )");
-    tool::AddVectorSink("multiplied_ints", &graph_config_, &output_packets_);
+    tool::AddVectorSink("multiplied_ints", &graph_config, &output_packets_);
+    MP_ASSERT_OK(graph_.Initialize(graph_config));
+    MP_ASSERT_OK(graph_.StartRun({}));
   }
 
-  CalculatorGraphConfig graph_config_;
+  void SendPackets(Timestamp timestamp, int multiplier, std::vector<int> ints) {
+    MP_ASSERT_OK(graph_.AddPacketToInputStream(
+        "ints", MakePacket<std::vector<int>>(std::move(ints)).At(timestamp)));
+    MP_ASSERT_OK(graph_.AddPacketToInputStream(
+        "multiplier", MakePacket<int>(multiplier).At(timestamp)));
+  }
+
+  void SendMultiplier(Timestamp timestamp, int multiplier) {
+    MP_ASSERT_OK(graph_.AddPacketToInputStream(
+        "multiplier", MakePacket<int>(multiplier).At(timestamp)));
+  }
+
+  CalculatorGraph graph_;
   std::vector<Packet> output_packets_;
 };
 
-TEST_F(BeginEndLoopCalculatorGraphWithClonedInputsTest, SingleEmptyVector) {
-  CalculatorGraph graph;
-  MP_EXPECT_OK(graph.Initialize(graph_config_));
-  MP_EXPECT_OK(graph.StartRun({}));
-  auto input_vector = absl::make_unique<std::vector<int>>();
+TEST_F(BeginEndLoopCalculatorGraphWithClonedInputsTest,
+       InputStreamForIterableIsEmpty) {
   Timestamp input_timestamp = Timestamp(42);
-  MP_ASSERT_OK(graph.AddPacketToInputStream(
-      "ints", Adopt(input_vector.release()).At(input_timestamp)));
-  auto multiplier = absl::make_unique<int>(2);
-  MP_ASSERT_OK(graph.AddPacketToInputStream(
-      "multiplier", Adopt(multiplier.release()).At(input_timestamp)));
-  MP_ASSERT_OK(graph.WaitUntilIdle());
+  SendMultiplier(input_timestamp, /*multiplier=*/2);
+  MP_ASSERT_OK(graph_.WaitUntilIdle());
+
+  // EndLoopCalc will forward the timestamp bound because there are no packets
+  // to process.
+  ASSERT_EQ(0, output_packets_.size());
+
+  MP_ASSERT_OK(graph_.CloseAllPacketSources());
+  MP_ASSERT_OK(graph_.WaitUntilDone());
+}
+
+TEST_F(BeginEndLoopCalculatorGraphWithClonedInputsTest, SingleEmptyVector) {
+  SendPackets(Timestamp(0), /*multiplier=*/2, /*ints=*/{});
+  MP_ASSERT_OK(graph_.WaitUntilIdle());
 
   // EndLoopCalc will forward the timestamp bound because there are no elements
   // in collection to output.
-  ASSERT_EQ(0, output_packets_.size());
+  EXPECT_TRUE(output_packets_.empty());
 
-  MP_ASSERT_OK(graph.CloseAllPacketSources());
-  MP_ASSERT_OK(graph.WaitUntilDone());
+  MP_ASSERT_OK(graph_.CloseAllPacketSources());
+  MP_ASSERT_OK(graph_.WaitUntilDone());
 }
 
 TEST_F(BeginEndLoopCalculatorGraphWithClonedInputsTest, SingleNonEmptyVector) {
-  CalculatorGraph graph;
-  MP_EXPECT_OK(graph.Initialize(graph_config_));
-  MP_EXPECT_OK(graph.StartRun({}));
-  auto input_vector = absl::make_unique<std::vector<int>>();
-  input_vector->emplace_back(0);
-  input_vector->emplace_back(1);
-  input_vector->emplace_back(2);
   Timestamp input_timestamp = Timestamp(42);
-  MP_ASSERT_OK(graph.AddPacketToInputStream(
-      "ints", Adopt(input_vector.release()).At(input_timestamp)));
-  auto multiplier = absl::make_unique<int>(2);
-  MP_ASSERT_OK(graph.AddPacketToInputStream(
-      "multiplier", Adopt(multiplier.release()).At(input_timestamp)));
-  MP_ASSERT_OK(graph.WaitUntilIdle());
+  SendPackets(input_timestamp, /*multiplier=*/2, /*ints=*/{0, 1, 2});
+  MP_ASSERT_OK(graph_.WaitUntilIdle());
 
-  ASSERT_EQ(1, output_packets_.size());
-  EXPECT_EQ(input_timestamp, output_packets_[0].Timestamp());
-  std::vector<int> expected_output_vector = {0, 2, 4};
-  EXPECT_EQ(expected_output_vector, output_packets_[0].Get<std::vector<int>>());
+  EXPECT_THAT(output_packets_,
+              testing::ElementsAre(
+                  PacketOfIntsEq(input_timestamp, std::vector<int>{0, 2, 4})));
 
-  MP_ASSERT_OK(graph.CloseAllPacketSources());
-  MP_ASSERT_OK(graph.WaitUntilDone());
+  MP_ASSERT_OK(graph_.CloseAllPacketSources());
+  MP_ASSERT_OK(graph_.WaitUntilDone());
 }
 
 TEST_F(BeginEndLoopCalculatorGraphWithClonedInputsTest, MultipleVectors) {
-  CalculatorGraph graph;
-  MP_EXPECT_OK(graph.Initialize(graph_config_));
-  MP_EXPECT_OK(graph.StartRun({}));
-
-  auto input_vector0 = absl::make_unique<std::vector<int>>();
-  input_vector0->emplace_back(0);
-  input_vector0->emplace_back(1);
   Timestamp input_timestamp0 = Timestamp(42);
-  MP_ASSERT_OK(graph.AddPacketToInputStream(
-      "ints", Adopt(input_vector0.release()).At(input_timestamp0)));
-  auto multiplier0 = absl::make_unique<int>(2);
-  MP_ASSERT_OK(graph.AddPacketToInputStream(
-      "multiplier", Adopt(multiplier0.release()).At(input_timestamp0)));
+  SendPackets(input_timestamp0, /*multiplier=*/2, /*ints=*/{0, 1});
 
-  auto input_vector1 = absl::make_unique<std::vector<int>>();
   Timestamp input_timestamp1 = Timestamp(43);
-  MP_ASSERT_OK(graph.AddPacketToInputStream(
-      "ints", Adopt(input_vector1.release()).At(input_timestamp1)));
-  auto multiplier1 = absl::make_unique<int>(2);
-  MP_ASSERT_OK(graph.AddPacketToInputStream(
-      "multiplier", Adopt(multiplier1.release()).At(input_timestamp1)));
+  SendPackets(input_timestamp1, /*multiplier=*/2, /*ints=*/{});
 
-  auto input_vector2 = absl::make_unique<std::vector<int>>();
-  input_vector2->emplace_back(2);
-  input_vector2->emplace_back(3);
   Timestamp input_timestamp2 = Timestamp(44);
-  MP_ASSERT_OK(graph.AddPacketToInputStream(
-      "ints", Adopt(input_vector2.release()).At(input_timestamp2)));
-  auto multiplier2 = absl::make_unique<int>(3);
-  MP_ASSERT_OK(graph.AddPacketToInputStream(
-      "multiplier", Adopt(multiplier2.release()).At(input_timestamp2)));
+  SendPackets(input_timestamp2, /*multiplier=*/3, /*ints=*/{2, 3});
 
-  MP_ASSERT_OK(graph.CloseAllPacketSources());
-  MP_ASSERT_OK(graph.WaitUntilDone());
-
-  ASSERT_EQ(2, output_packets_.size());
-
-  EXPECT_EQ(input_timestamp0, output_packets_[0].Timestamp());
-  std::vector<int> expected_output_vector0 = {0, 2};
-  EXPECT_EQ(expected_output_vector0,
-            output_packets_[0].Get<std::vector<int>>());
+  MP_ASSERT_OK(graph_.CloseAllPacketSources());
+  MP_ASSERT_OK(graph_.WaitUntilDone());
 
   // At input_timestamp1, EndLoopCalc will forward timestamp bound as there are
   // no elements in vector to process.
-
-  EXPECT_EQ(input_timestamp2, output_packets_[1].Timestamp());
-  std::vector<int> expected_output_vector2 = {6, 9};
-  EXPECT_EQ(expected_output_vector2,
-            output_packets_[1].Get<std::vector<int>>());
+  EXPECT_THAT(output_packets_,
+              testing::ElementsAre(
+                  PacketOfIntsEq(input_timestamp0, std::vector<int>{0, 2}),
+                  PacketOfIntsEq(input_timestamp2, std::vector<int>{6, 9})));
 }
 
 }  // namespace

mediapipe/calculators/core/begin_loop_calculator.cc

@@ -16,7 +16,9 @@
 
 #include <vector>
 
+#include "mediapipe/framework/formats/detection.pb.h"
 #include "mediapipe/framework/formats/landmark.pb.h"
+#include "mediapipe/framework/formats/matrix.h"
 #include "mediapipe/framework/formats/rect.pb.h"
 
 namespace mediapipe {
@@ -31,4 +33,13 @@ typedef BeginLoopCalculator<std::vector<::mediapipe::NormalizedRect>>
     BeginLoopNormalizedRectCalculator;
 REGISTER_CALCULATOR(BeginLoopNormalizedRectCalculator);
 
+// A calculator to process std::vector<Detection>.
+typedef BeginLoopCalculator<std::vector<::mediapipe::Detection>>
+    BeginLoopDetectionCalculator;
+REGISTER_CALCULATOR(BeginLoopDetectionCalculator);
+
+// A calculator to process std::vector<Matrix>.
+typedef BeginLoopCalculator<std::vector<Matrix>> BeginLoopMatrixCalculator;
+REGISTER_CALCULATOR(BeginLoopMatrixCalculator);
+
 }  // namespace mediapipe

mediapipe/calculators/core/begin_loop_calculator.h

@@ -52,20 +52,28 @@ namespace mediapipe {
 //   output_stream: "OUTPUT:aggregated_result"     # IterableU @ext_ts
 // }
 //
-// BeginLoopCalculator accepts an optional input stream tagged with "TICK"
-// which if non-empty, wakes up the calculator and calls
-// BeginLoopCalculator::Process(). Input streams tagged with "CLONE" are cloned
-// to the corresponding output streams at loop timestamps. This ensures that a
-// MediaPipe graph or sub-graph can run multiple times, once per element in the
-// "ITERABLE" for each pakcet clone of the packets in the "CLONE" input streams.
+// Input streams tagged with "CLONE" are cloned to the corresponding output
+// streams at loop timestamps. This ensures that a MediaPipe graph or sub-graph
+// can run multiple times, once per element in the "ITERABLE" for each pakcet
+// clone of the packets in the "CLONE" input streams.
 template <typename IterableT>
 class BeginLoopCalculator : public CalculatorBase {
   using ItemT = typename IterableT::value_type;
 
  public:
   static ::mediapipe::Status GetContract(CalculatorContract* cc) {
+    // The below enables processing of timestamp bound updates, and that enables
+    // correct timestamp propagation by the companion EndLoopCalculator.
+    //
+    // For instance, Process() function will be still invoked even if upstream
+    // calculator has updated timestamp bound for ITERABLE input instead of
+    // providing actual value.
+    cc->SetProcessTimestampBounds(true);
+
     // A non-empty packet in the optional "TICK" input stream wakes up the
     // calculator.
+    // DEPRECATED as timestamp bound updates are processed by default in this
+    // calculator.
     if (cc->Inputs().HasTag("TICK")) {
       cc->Inputs().Tag("TICK").SetAny();
     }

mediapipe/calculators/core/clip_vector_size_calculator.cc

@@ -17,6 +17,7 @@
 #include <vector>
 
 #include "mediapipe/framework/calculator_framework.h"
+#include "mediapipe/framework/formats/detection.pb.h"
 #include "mediapipe/framework/formats/rect.pb.h"
 
 namespace mediapipe {
@@ -25,4 +26,8 @@ typedef ClipVectorSizeCalculator<::mediapipe::NormalizedRect>
     ClipNormalizedRectVectorSizeCalculator;
 REGISTER_CALCULATOR(ClipNormalizedRectVectorSizeCalculator);
 
+typedef ClipVectorSizeCalculator<::mediapipe::Detection>
+    ClipDetectionVectorSizeCalculator;
+REGISTER_CALCULATOR(ClipDetectionVectorSizeCalculator);
+
 }  // namespace mediapipe

mediapipe/calculators/core/end_loop_calculator.cc

@@ -16,9 +16,11 @@
 
 #include <vector>
 
+#include "mediapipe/framework/formats/classification.pb.h"
 #include "mediapipe/framework/formats/landmark.pb.h"
 #include "mediapipe/framework/formats/rect.pb.h"
 #include "mediapipe/util/render_data.pb.h"
+#include "tensorflow/lite/interpreter.h"
 
 namespace mediapipe {
 
@@ -37,4 +39,11 @@ typedef EndLoopCalculator<std::vector<::mediapipe::RenderData>>
     EndLoopRenderDataCalculator;
 REGISTER_CALCULATOR(EndLoopRenderDataCalculator);
 
+typedef EndLoopCalculator<std::vector<::mediapipe::ClassificationList>>
+    EndLoopClassificationListCalculator;
+REGISTER_CALCULATOR(EndLoopClassificationListCalculator);
+
+typedef EndLoopCalculator<std::vector<TfLiteTensor>> EndLoopTensorCalculator;
+REGISTER_CALCULATOR(EndLoopTensorCalculator);
+
 }  // namespace mediapipe

mediapipe/calculators/core/previous_loopback_calculator.cc

@@ -25,13 +25,17 @@ namespace mediapipe {
 // together with some previous output.
 //
 // For the first packet that arrives on the MAIN input, the timestamp bound is
-// advanced on the output. Downstream calculators will see this as an empty
+// advanced on the PREV_LOOP. Downstream calculators will see this as an empty
 // packet. This way they are not kept waiting for the previous output, which
 // for the first iteration does not exist.
 //
-// Thereafter, each packet received on MAIN is matched with a packet received
-// on LOOP; the LOOP packet's timestamp is changed to that of the MAIN packet,
-// and it is output on PREV_LOOP.
+// Thereafter,
+// - Each non-empty MAIN packet results in:
+//   a) a PREV_LOOP packet with contents of the LOOP packet received at the
+//      timestamp of the previous non-empty MAIN packet
+//   b) or in a PREV_LOOP timestamp bound update if the LOOP packet was empty.
+// - Each empty MAIN packet indicating timestamp bound update results in a
+//   PREV_LOOP timestamp bound update.
 //
 // Example config:
 // node {
@@ -56,83 +60,115 @@ class PreviousLoopbackCalculator : public CalculatorBase {
     // TODO: an optional PREV_TIMESTAMP output could be added to
     // carry the original timestamp of the packet on PREV_LOOP.
     cc->SetInputStreamHandler("ImmediateInputStreamHandler");
+    // Process() function is invoked in response to MAIN/LOOP stream timestamp
+    // bound updates.
+    cc->SetProcessTimestampBounds(true);
     return ::mediapipe::OkStatus();
   }
 
   ::mediapipe::Status Open(CalculatorContext* cc) final {
     main_id_ = cc->Inputs().GetId("MAIN", 0);
     loop_id_ = cc->Inputs().GetId("LOOP", 0);
-    loop_out_id_ = cc->Outputs().GetId("PREV_LOOP", 0);
+    prev_loop_id_ = cc->Outputs().GetId("PREV_LOOP", 0);
     cc->Outputs()
-        .Get(loop_out_id_)
+        .Get(prev_loop_id_)
         .SetHeader(cc->Inputs().Get(loop_id_).Header());
-
-    // Use an empty packet for the first round, since there is no previous
-    // output.
-    loopback_packets_.push_back({});
-
     return ::mediapipe::OkStatus();
   }
 
   ::mediapipe::Status Process(CalculatorContext* cc) final {
-    Packet& main_packet = cc->Inputs().Get(main_id_).Value();
-    if (!main_packet.IsEmpty()) {
-      main_ts_.push_back(main_packet.Timestamp());
-    }
-    Packet& loopback_packet = cc->Inputs().Get(loop_id_).Value();
-    if (!loopback_packet.IsEmpty()) {
-      loopback_packets_.push_back(loopback_packet);
-      while (!main_ts_.empty() &&
-             main_ts_.front() <= loopback_packets_.front().Timestamp()) {
-        main_ts_.pop_front();
-      }
-    }
-    auto& loop_out = cc->Outputs().Get(loop_out_id_);
+    // Non-empty packets and empty packets indicating timestamp bound updates
+    // are guaranteed to have timestamps greater than timestamps of previous
+    // packets within the same stream. Calculator tracks and operates on such
+    // packets.
 
-    while (!main_ts_.empty() && !loopback_packets_.empty()) {
-      Timestamp main_timestamp = main_ts_.front();
-      main_ts_.pop_front();
-      Packet previous_loopback = loopback_packets_.front().At(main_timestamp);
-      loopback_packets_.pop_front();
-
-      if (previous_loopback.IsEmpty()) {
-        // TODO: SetCompleteTimestampBound would be more useful.
-        loop_out.SetNextTimestampBound(main_timestamp + 1);
+    const Packet& main_packet = cc->Inputs().Get(main_id_).Value();
+    if (prev_main_ts_ < main_packet.Timestamp()) {
+      Timestamp loop_timestamp;
+      if (!main_packet.IsEmpty()) {
+        loop_timestamp = prev_non_empty_main_ts_;
+        prev_non_empty_main_ts_ = main_packet.Timestamp();
       } else {
-        loop_out.AddPacket(std::move(previous_loopback));
+        // Calculator advances PREV_LOOP timestamp bound in response to empty
+        // MAIN packet, hence not caring about corresponding loop packet.
+        loop_timestamp = Timestamp::Unset();
+      }
+      main_packet_specs_.push_back({.timestamp = main_packet.Timestamp(),
+                                    .loop_timestamp = loop_timestamp});
+      prev_main_ts_ = main_packet.Timestamp();
+    }
+
+    const Packet& loop_packet = cc->Inputs().Get(loop_id_).Value();
+    if (prev_loop_ts_ < loop_packet.Timestamp()) {
+      loop_packets_.push_back(loop_packet);
+      prev_loop_ts_ = loop_packet.Timestamp();
+    }
+
+    auto& prev_loop = cc->Outputs().Get(prev_loop_id_);
+    while (!main_packet_specs_.empty() && !loop_packets_.empty()) {
+      // The earliest MAIN packet.
+      const MainPacketSpec& main_spec = main_packet_specs_.front();
+      // The earliest LOOP packet.
+      const Packet& loop_candidate = loop_packets_.front();
+      // Match LOOP and MAIN packets.
+      if (main_spec.loop_timestamp < loop_candidate.Timestamp()) {
+        // No LOOP packet can match the MAIN packet under review.
+        prev_loop.SetNextTimestampBound(main_spec.timestamp + 1);
+        main_packet_specs_.pop_front();
+      } else if (main_spec.loop_timestamp > loop_candidate.Timestamp()) {
+        // No MAIN packet can match the LOOP packet under review.
+        loop_packets_.pop_front();
+      } else {
+        // Exact match found.
+        if (loop_candidate.IsEmpty()) {
+          // However, LOOP packet is empty.
+          prev_loop.SetNextTimestampBound(main_spec.timestamp + 1);
+        } else {
+          prev_loop.AddPacket(loop_candidate.At(main_spec.timestamp));
+        }
+        loop_packets_.pop_front();
+        main_packet_specs_.pop_front();
       }
     }
 
-    // In case of an empty loopback input, the next timestamp bound for
-    // loopback input is the loopback timestamp + 1.  The next timestamp bound
-    // for output is set and the main_ts_ vector is truncated accordingly.
-    if (loopback_packet.IsEmpty() &&
-        loopback_packet.Timestamp() != Timestamp::Unstarted()) {
-      Timestamp loopback_bound =
-          loopback_packet.Timestamp().NextAllowedInStream();
-      while (!main_ts_.empty() && main_ts_.front() <= loopback_bound) {
-        main_ts_.pop_front();
-      }
-      if (main_ts_.empty()) {
-        loop_out.SetNextTimestampBound(loopback_bound.NextAllowedInStream());
-      }
-    }
-    if (!main_ts_.empty()) {
-      loop_out.SetNextTimestampBound(main_ts_.front());
-    }
-    if (cc->Inputs().Get(main_id_).IsDone() && main_ts_.empty()) {
-      loop_out.Close();
+    if (main_packet_specs_.empty() && cc->Inputs().Get(main_id_).IsDone()) {
+      prev_loop.Close();
     }
     return ::mediapipe::OkStatus();
   }
 
  private:
+  struct MainPacketSpec {
+    Timestamp timestamp;
+    // Expected timestamp of the packet from LOOP stream that corresponds to the
+    // packet from MAIN stream descirbed by this spec.
+    Timestamp loop_timestamp;
+  };
+
   CollectionItemId main_id_;
   CollectionItemId loop_id_;
-  CollectionItemId loop_out_id_;
+  CollectionItemId prev_loop_id_;
 
-  std::deque<Timestamp> main_ts_;
-  std::deque<Packet> loopback_packets_;
+  // Contains specs for MAIN packets which only can be:
+  // - non-empty packets
+  // - empty packets indicating timestamp bound updates
+  //
+  // Sorted according to packet timestamps.
+  std::deque<MainPacketSpec> main_packet_specs_;
+  Timestamp prev_main_ts_ = Timestamp::Unstarted();
+  Timestamp prev_non_empty_main_ts_ = Timestamp::Unstarted();
+
+  // Contains LOOP packets which only can be:
+  // - the very first empty packet
+  // - non empty packets
+  // - empty packets indicating timestamp bound updates
+  //
+  // Sorted according to packet timestamps.
+  std::deque<Packet> loop_packets_;
+  // Using "Timestamp::Unset" instead of "Timestamp::Unstarted" in order to
+  // allow addition of the very first empty packet (which doesn't indicate
+  // timestamp bound change necessarily).
+  Timestamp prev_loop_ts_ = Timestamp::Unset();
 };
 REGISTER_CALCULATOR(PreviousLoopbackCalculator);
 

mediapipe/calculators/core/previous_loopback_calculator_test.cc

@@ -13,6 +13,7 @@
 // limitations under the License.
 
 #include <algorithm>
+#include <functional>
 #include <memory>
 #include <string>
 #include <vector>
@@ -25,12 +26,17 @@
 #include "mediapipe/framework/port/gtest.h"
 #include "mediapipe/framework/port/integral_types.h"
 #include "mediapipe/framework/port/parse_text_proto.h"
+#include "mediapipe/framework/port/status.h"
 #include "mediapipe/framework/port/status_matchers.h"
 #include "mediapipe/framework/timestamp.h"
 #include "mediapipe/framework/tool/sink.h"
 
 namespace mediapipe {
 
+using ::testing::ElementsAre;
+using ::testing::Eq;
+using ::testing::Pair;
+using ::testing::Value;
 namespace {
 
 // Returns the timestamp values for a vector of Packets.
@@ -43,6 +49,23 @@ std::vector<int64> TimestampValues(const std::vector<Packet>& packets) {
   return result;
 }
 
+MATCHER(EmptyPacket, negation ? "isn't empty" : "is empty") {
+  if (arg.IsEmpty()) {
+    return true;
+  }
+  return false;
+}
+
+MATCHER_P(IntPacket, value, "") {
+  return Value(arg.template Get<int>(), Eq(value));
+}
+
+MATCHER_P2(PairPacket, timestamp, pair, "") {
+  Timestamp actual_timestamp = arg.Timestamp();
+  const auto& actual_pair = arg.template Get<std::pair<Packet, Packet>>();
+  return Value(actual_timestamp, Eq(timestamp)) && Value(actual_pair, pair);
+}
+
 TEST(PreviousLoopbackCalculator, CorrectTimestamps) {
   std::vector<Packet> in_prev;
   CalculatorGraphConfig graph_config_ =
@@ -81,32 +104,30 @@ TEST(PreviousLoopbackCalculator, CorrectTimestamps) {
     MP_EXPECT_OK(graph_.AddPacketToInputStream(
         input_name, MakePacket<int>(n).At(Timestamp(n))));
   };
-  auto pair_values = [](const Packet& packet) {
-    auto pair = packet.Get<std::pair<Packet, Packet>>();
-    int first = pair.first.IsEmpty() ? -1 : pair.first.Get<int>();
-    int second = pair.second.IsEmpty() ? -1 : pair.second.Get<int>();
-    return std::make_pair(first, second);
-  };
 
   send_packet("in", 1);
   MP_EXPECT_OK(graph_.WaitUntilIdle());
-  EXPECT_EQ(TimestampValues(in_prev), (std::vector<int64>{1}));
-  EXPECT_EQ(pair_values(in_prev.back()), std::make_pair(1, -1));
+  EXPECT_THAT(TimestampValues(in_prev), ElementsAre(1));
+  EXPECT_THAT(in_prev.back(),
+              PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())));
 
   send_packet("in", 2);
   MP_EXPECT_OK(graph_.WaitUntilIdle());
-  EXPECT_EQ(TimestampValues(in_prev), (std::vector<int64>{1, 2}));
-  EXPECT_EQ(pair_values(in_prev.back()), std::make_pair(2, 1));
+  EXPECT_THAT(TimestampValues(in_prev), ElementsAre(1, 2));
+  EXPECT_THAT(in_prev.back(),
+              PairPacket(Timestamp(2), Pair(IntPacket(2), IntPacket(1))));
 
   send_packet("in", 5);
   MP_EXPECT_OK(graph_.WaitUntilIdle());
-  EXPECT_EQ(TimestampValues(in_prev), (std::vector<int64>{1, 2, 5}));
-  EXPECT_EQ(pair_values(in_prev.back()), std::make_pair(5, 2));
+  EXPECT_THAT(TimestampValues(in_prev), ElementsAre(1, 2, 5));
+  EXPECT_THAT(in_prev.back(),
+              PairPacket(Timestamp(5), Pair(IntPacket(5), IntPacket(2))));
 
   send_packet("in", 15);
   MP_EXPECT_OK(graph_.WaitUntilIdle());
-  EXPECT_EQ(TimestampValues(in_prev), (std::vector<int64>{1, 2, 5, 15}));
-  EXPECT_EQ(pair_values(in_prev.back()), std::make_pair(15, 5));
+  EXPECT_THAT(TimestampValues(in_prev), ElementsAre(1, 2, 5, 15));
+  EXPECT_THAT(in_prev.back(),
+              PairPacket(Timestamp(15), Pair(IntPacket(15), IntPacket(5))));
 
   MP_EXPECT_OK(graph_.CloseAllInputStreams());
   MP_EXPECT_OK(graph_.WaitUntilDone());
@@ -185,24 +206,24 @@ TEST(PreviousLoopbackCalculator, ClosesCorrectly) {
 
   send_packet("in", 1);
   MP_EXPECT_OK(graph_.WaitUntilIdle());
-  EXPECT_EQ(TimestampValues(outputs), (std::vector<int64>{1}));
+  EXPECT_THAT(TimestampValues(outputs), ElementsAre(1));
 
   send_packet("in", 2);
   MP_EXPECT_OK(graph_.WaitUntilIdle());
-  EXPECT_EQ(TimestampValues(outputs), (std::vector<int64>{1, 2}));
+  EXPECT_THAT(TimestampValues(outputs), ElementsAre(1, 2));
 
   send_packet("in", 5);
   MP_EXPECT_OK(graph_.WaitUntilIdle());
-  EXPECT_EQ(TimestampValues(outputs), (std::vector<int64>{1, 2, 5}));
+  EXPECT_THAT(TimestampValues(outputs), ElementsAre(1, 2, 5));
 
   send_packet("in", 15);
   MP_EXPECT_OK(graph_.WaitUntilIdle());
-  EXPECT_EQ(TimestampValues(outputs), (std::vector<int64>{1, 2, 5, 15}));
+  EXPECT_THAT(TimestampValues(outputs), ElementsAre(1, 2, 5, 15));
 
   MP_EXPECT_OK(graph_.CloseAllInputStreams());
   MP_EXPECT_OK(graph_.WaitUntilIdle());
-  EXPECT_EQ(TimestampValues(outputs),
-            (std::vector<int64>{1, 2, 5, 15, Timestamp::Max().Value()}));
+  EXPECT_THAT(TimestampValues(outputs),
+              ElementsAre(1, 2, 5, 15, Timestamp::Max().Value()));
 
   MP_EXPECT_OK(graph_.WaitUntilDone());
 }
@@ -247,16 +268,12 @@ TEST(PreviousLoopbackCalculator, EmptyLoopForever) {
         input_name, MakePacket<int>(n).At(Timestamp(n))));
   };
 
-  send_packet("in", 0);
-  MP_EXPECT_OK(graph_.WaitUntilIdle());
-  EXPECT_EQ(TimestampValues(outputs), (std::vector<int64>{0}));
-
-  for (int main_ts = 1; main_ts < 50; ++main_ts) {
+  for (int main_ts = 0; main_ts < 50; ++main_ts) {
     send_packet("in", main_ts);
     MP_EXPECT_OK(graph_.WaitUntilIdle());
     std::vector<int64> ts_values = TimestampValues(outputs);
     EXPECT_EQ(ts_values.size(), main_ts + 1);
-    for (int j = 0; j < main_ts; ++j) {
+    for (int j = 0; j < main_ts + 1; ++j) {
       EXPECT_EQ(ts_values[j], j);
     }
   }
@@ -266,5 +283,487 @@ TEST(PreviousLoopbackCalculator, EmptyLoopForever) {
   MP_EXPECT_OK(graph_.WaitUntilDone());
 }
 
+class PreviousLoopbackCalculatorProcessingTimestampsTest
+    : public testing::Test {
+ protected:
+  void SetUp() override {
+    CalculatorGraphConfig graph_config =
+        ParseTextProtoOrDie<CalculatorGraphConfig>(R"(
+          input_stream: 'input'
+          input_stream: 'force_main_empty'
+          input_stream: 'force_loop_empty'
+          # Used to indicate "main" timestamp bound updates.
+          node {
+            calculator: 'GateCalculator'
+            input_stream: 'input'
+            input_stream: 'DISALLOW:force_main_empty'
+            output_stream: 'main'
+          }
+          node {
+            calculator: 'PreviousLoopbackCalculator'
+            input_stream: 'MAIN:main'
+            input_stream: 'LOOP:loop'
+            input_stream_info: { tag_index: 'LOOP' back_edge: true }
+            output_stream: 'PREV_LOOP:prev_loop'
+          }
+          node {
+            calculator: 'PassThroughCalculator'
+            input_stream: 'input'
+            input_stream: 'prev_loop'
+            output_stream: 'passed_through_input'
+            output_stream: 'passed_through_prev_loop'
+          }
+          # Used to indicate "loop" timestamp bound updates.
+          node {
+            calculator: 'GateCalculator'
+            input_stream: 'input'
+            input_stream: 'DISALLOW:force_loop_empty'
+            output_stream: 'loop'
+          }
+          node {
+            calculator: 'MakePairCalculator'
+            input_stream: 'passed_through_input'
+            input_stream: 'passed_through_prev_loop'
+            output_stream: 'passed_through_input_and_prev_loop'
+          }
+        )");
+    tool::AddVectorSink("passed_through_input_and_prev_loop", &graph_config,
+                        &output_packets_);
+    MP_ASSERT_OK(graph_.Initialize(graph_config, {}));
+    MP_ASSERT_OK(graph_.StartRun({}));
+  }
+
+  void SendPackets(int timestamp, int input, bool force_main_empty,
+                   bool force_loop_empty) {
+    MP_ASSERT_OK(graph_.AddPacketToInputStream(
+        "input", MakePacket<int>(input).At(Timestamp(timestamp))));
+    MP_ASSERT_OK(graph_.AddPacketToInputStream(
+        "force_main_empty",
+        MakePacket<bool>(force_main_empty).At(Timestamp(timestamp))));
+    MP_ASSERT_OK(graph_.AddPacketToInputStream(
+        "force_loop_empty",
+        MakePacket<bool>(force_loop_empty).At(Timestamp(timestamp))));
+  }
+
+  CalculatorGraph graph_;
+  std::vector<Packet> output_packets_;
+};
+
+TEST_F(PreviousLoopbackCalculatorProcessingTimestampsTest,
+       MultiplePacketsEmptyMainNonEmptyLoop) {
+  SendPackets(/*timestamp=*/1, /*input=*/1, /*force_main_empty=*/true,
+              /*force_loop_empty=*/false);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket()))));
+
+  SendPackets(/*timestamp=*/2, /*input=*/2, /*force_main_empty=*/true,
+              /*force_loop_empty=*/false);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())),
+                  PairPacket(Timestamp(2), Pair(IntPacket(2), EmptyPacket()))));
+
+  SendPackets(/*timestamp=*/3, /*input=*/3, /*force_main_empty=*/true,
+              /*force_loop_empty=*/false);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())),
+                  PairPacket(Timestamp(2), Pair(IntPacket(2), EmptyPacket())),
+                  PairPacket(Timestamp(3), Pair(IntPacket(3), EmptyPacket()))));
+
+  SendPackets(/*timestamp=*/5, /*input=*/5, /*force_main_empty=*/true,
+              /*force_loop_empty=*/false);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())),
+                  PairPacket(Timestamp(2), Pair(IntPacket(2), EmptyPacket())),
+                  PairPacket(Timestamp(3), Pair(IntPacket(3), EmptyPacket())),
+                  PairPacket(Timestamp(5), Pair(IntPacket(5), EmptyPacket()))));
+
+  SendPackets(/*timestamp=*/15, /*input=*/15,
+              /*force_main_empty=*/true,
+              /*force_loop_empty=*/false);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(
+          PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())),
+          PairPacket(Timestamp(2), Pair(IntPacket(2), EmptyPacket())),
+          PairPacket(Timestamp(3), Pair(IntPacket(3), EmptyPacket())),
+          PairPacket(Timestamp(5), Pair(IntPacket(5), EmptyPacket())),
+          PairPacket(Timestamp(15), Pair(IntPacket(15), EmptyPacket()))));
+
+  MP_EXPECT_OK(graph_.CloseAllInputStreams());
+  MP_EXPECT_OK(graph_.WaitUntilDone());
+}
+
+TEST_F(PreviousLoopbackCalculatorProcessingTimestampsTest,
+       MultiplePacketsNonEmptyMainEmptyLoop) {
+  SendPackets(/*timestamp=*/1, /*input=*/1,
+              /*force_main_empty=*/false,
+              /*force_loop_empty=*/true);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket()))));
+
+  SendPackets(/*timestamp=*/2, /*input=*/2,
+              /*force_main_empty=*/false,
+              /*force_loop_empty=*/true);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())),
+                  PairPacket(Timestamp(2), Pair(IntPacket(2), EmptyPacket()))));
+
+  SendPackets(/*timestamp=*/3, /*input=*/3,
+              /*force_main_empty=*/false,
+              /*force_loop_empty=*/true);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())),
+                  PairPacket(Timestamp(2), Pair(IntPacket(2), EmptyPacket())),
+                  PairPacket(Timestamp(3), Pair(IntPacket(3), EmptyPacket()))));
+
+  SendPackets(/*timestamp=*/5, /*input=*/5,
+              /*force_main_empty=*/false,
+              /*force_loop_empty=*/true);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())),
+                  PairPacket(Timestamp(2), Pair(IntPacket(2), EmptyPacket())),
+                  PairPacket(Timestamp(3), Pair(IntPacket(3), EmptyPacket())),
+                  PairPacket(Timestamp(5), Pair(IntPacket(5), EmptyPacket()))));
+
+  SendPackets(/*timestamp=*/15, /*input=*/15,
+              /*force_main_empty=*/false,
+              /*force_loop_empty=*/true);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(
+          PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())),
+          PairPacket(Timestamp(2), Pair(IntPacket(2), EmptyPacket())),
+          PairPacket(Timestamp(3), Pair(IntPacket(3), EmptyPacket())),
+          PairPacket(Timestamp(5), Pair(IntPacket(5), EmptyPacket())),
+          PairPacket(Timestamp(15), Pair(IntPacket(15), EmptyPacket()))));
+
+  MP_EXPECT_OK(graph_.CloseAllInputStreams());
+  MP_EXPECT_OK(graph_.WaitUntilDone());
+}
+
+TEST_F(PreviousLoopbackCalculatorProcessingTimestampsTest,
+       MultiplePacketsAlteringMainNonEmptyLoop) {
+  SendPackets(/*timestamp=*/1, /*input=*/1,
+              /*force_main_empty=*/false,
+              /*force_loop_empty=*/false);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket()))));
+
+  SendPackets(/*timestamp=*/2, /*input=*/2, /*force_main_empty=*/true,
+              /*force_loop_empty=*/false);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())),
+                  PairPacket(Timestamp(2), Pair(IntPacket(2), EmptyPacket()))));
+
+  SendPackets(/*timestamp=*/3, /*input=*/3,
+              /*force_main_empty=*/false,
+              /*force_loop_empty=*/false);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())),
+                  PairPacket(Timestamp(2), Pair(IntPacket(2), EmptyPacket())),
+                  PairPacket(Timestamp(3), Pair(IntPacket(3), IntPacket(1)))));
+
+  SendPackets(/*timestamp=*/5, /*input=*/5,
+              /*force_main_empty=*/false,
+              /*force_loop_empty=*/false);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())),
+                  PairPacket(Timestamp(2), Pair(IntPacket(2), EmptyPacket())),
+                  PairPacket(Timestamp(3), Pair(IntPacket(3), IntPacket(1))),
+                  PairPacket(Timestamp(5), Pair(IntPacket(5), IntPacket(3)))));
+
+  SendPackets(/*timestamp=*/15, /*input=*/15,
+              /*force_main_empty=*/true,
+              /*force_loop_empty=*/false);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(
+          PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())),
+          PairPacket(Timestamp(2), Pair(IntPacket(2), EmptyPacket())),
+          PairPacket(Timestamp(3), Pair(IntPacket(3), IntPacket(1))),
+          PairPacket(Timestamp(5), Pair(IntPacket(5), IntPacket(3))),
+          PairPacket(Timestamp(15), Pair(IntPacket(15), EmptyPacket()))));
+
+  MP_EXPECT_OK(graph_.CloseAllInputStreams());
+  MP_EXPECT_OK(graph_.WaitUntilDone());
+}
+
+TEST_F(PreviousLoopbackCalculatorProcessingTimestampsTest,
+       MultiplePacketsNonEmptyMainAlteringLoop) {
+  SendPackets(/*timestamp=*/1, /*input=*/1,
+              /*force_main_empty=*/false,
+              /*force_loop_empty=*/false);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket()))));
+
+  SendPackets(/*timestamp=*/2, /*input=*/2,
+              /*force_main_empty=*/false,
+              /*force_loop_empty=*/true);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())),
+                  PairPacket(Timestamp(2), Pair(IntPacket(2), IntPacket(1)))));
+
+  SendPackets(/*timestamp=*/3, /*input=*/3,
+              /*force_main_empty=*/false,
+              /*force_loop_empty=*/false);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())),
+                  PairPacket(Timestamp(2), Pair(IntPacket(2), IntPacket(1))),
+                  PairPacket(Timestamp(3), Pair(IntPacket(3), EmptyPacket()))));
+
+  SendPackets(/*timestamp=*/5, /*input=*/5,
+              /*force_main_empty=*/false,
+              /*force_loop_empty=*/true);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())),
+                  PairPacket(Timestamp(2), Pair(IntPacket(2), IntPacket(1))),
+                  PairPacket(Timestamp(3), Pair(IntPacket(3), EmptyPacket())),
+                  PairPacket(Timestamp(5), Pair(IntPacket(5), IntPacket(3)))));
+
+  SendPackets(/*timestamp=*/15, /*input=*/15,
+              /*force_main_empty=*/false,
+              /*force_loop_empty=*/false);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(
+          PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())),
+          PairPacket(Timestamp(2), Pair(IntPacket(2), IntPacket(1))),
+          PairPacket(Timestamp(3), Pair(IntPacket(3), EmptyPacket())),
+          PairPacket(Timestamp(5), Pair(IntPacket(5), IntPacket(3))),
+          PairPacket(Timestamp(15), Pair(IntPacket(15), EmptyPacket()))));
+
+  MP_EXPECT_OK(graph_.CloseAllInputStreams());
+  MP_EXPECT_OK(graph_.WaitUntilDone());
+}
+
+TEST_F(PreviousLoopbackCalculatorProcessingTimestampsTest,
+       MultiplePacketsCheckIfLastCorrectAlteringMainAlteringLoop) {
+  int num_packets = 1000;
+  for (int i = 0; i < num_packets; ++i) {
+    bool force_main_empty = i % 3 == 0 ? true : false;
+    bool force_loop_empty = i % 2 == 0 ? true : false;
+    SendPackets(/*timestamp=*/i + 1, /*input=*/i + 1, force_main_empty,
+                force_loop_empty);
+  }
+  SendPackets(/*timestamp=*/num_packets + 1,
+              /*input=*/num_packets + 1, /*force_main_empty=*/false,
+              /*force_loop_empty=*/false);
+  SendPackets(/*timestamp=*/num_packets + 2,
+              /*input=*/num_packets + 2, /*force_main_empty=*/false,
+              /*force_loop_empty=*/false);
+
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  ASSERT_FALSE(output_packets_.empty());
+  EXPECT_THAT(
+      output_packets_.back(),
+      PairPacket(Timestamp(num_packets + 2),
+                 Pair(IntPacket(num_packets + 2), IntPacket(num_packets + 1))));
+
+  MP_EXPECT_OK(graph_.CloseAllInputStreams());
+  MP_EXPECT_OK(graph_.WaitUntilDone());
+}
+
+// Similar to GateCalculator, but it doesn't propagate timestamp bound updates.
+class DroppingGateCalculator : public CalculatorBase {
+ public:
+  static ::mediapipe::Status GetContract(CalculatorContract* cc) {
+    cc->Inputs().Index(0).SetAny();
+    cc->Inputs().Tag("DISALLOW").Set<bool>();
+    cc->Outputs().Index(0).SetSameAs(&cc->Inputs().Index(0));
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Process(CalculatorContext* cc) final {
+    if (!cc->Inputs().Index(0).IsEmpty() &&
+        !cc->Inputs().Tag("DISALLOW").Get<bool>()) {
+      cc->Outputs().Index(0).AddPacket(cc->Inputs().Index(0).Value());
+    }
+    return ::mediapipe::OkStatus();
+  }
+};
+REGISTER_CALCULATOR(DroppingGateCalculator);
+
+// Tests PreviousLoopbackCalculator in cases when there are no "LOOP" timestamp
+// bound updates and non-empty packets for a while and the aforementioned start
+// to arrive at some point. So, "PREV_LOOP" is delayed for a couple of inputs.
+class PreviousLoopbackCalculatorDelayBehaviorTest : public testing::Test {
+ protected:
+  void SetUp() override {
+    CalculatorGraphConfig graph_config =
+        ParseTextProtoOrDie<CalculatorGraphConfig>(R"(
+          input_stream: 'input'
+          # Drops "loop" when set to "true", delaying output of prev_loop, hence
+          # delaying output of the graph.
+          input_stream: 'delay_next_output'
+          node {
+            calculator: 'PreviousLoopbackCalculator'
+            input_stream: 'MAIN:input'
+            input_stream: 'LOOP:loop'
+            input_stream_info: { tag_index: 'LOOP' back_edge: true }
+            output_stream: 'PREV_LOOP:prev_loop'
+          }
+          node {
+            calculator: 'PassThroughCalculator'
+            input_stream: 'input'
+            input_stream: 'prev_loop'
+            output_stream: 'passed_through_input'
+            output_stream: 'passed_through_prev_loop'
+          }
+          node {
+            calculator: 'DroppingGateCalculator'
+            input_stream: 'input'
+            input_stream: 'DISALLOW:delay_next_output'
+            output_stream: 'loop'
+          }
+          node {
+            calculator: 'MakePairCalculator'
+            input_stream: 'passed_through_input'
+            input_stream: 'passed_through_prev_loop'
+            output_stream: 'passed_through_input_and_prev_loop'
+          }
+        )");
+    tool::AddVectorSink("passed_through_input_and_prev_loop", &graph_config,
+                        &output_packets_);
+    MP_ASSERT_OK(graph_.Initialize(graph_config, {}));
+    MP_ASSERT_OK(graph_.StartRun({}));
+  }
+
+  void SendPackets(int timestamp, int input, bool delay_next_output) {
+    MP_ASSERT_OK(graph_.AddPacketToInputStream(
+        "input", MakePacket<int>(input).At(Timestamp(timestamp))));
+    MP_ASSERT_OK(graph_.AddPacketToInputStream(
+        "delay_next_output",
+        MakePacket<bool>(delay_next_output).At(Timestamp(timestamp))));
+  }
+
+  CalculatorGraph graph_;
+  std::vector<Packet> output_packets_;
+};
+
+TEST_F(PreviousLoopbackCalculatorDelayBehaviorTest, MultipleDelayedOutputs) {
+  SendPackets(/*timestamp=*/1, /*input=*/1, /*delay_next_output=*/true);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket()))));
+
+  SendPackets(/*timestamp=*/2, /*input=*/2, /*delay_next_output=*/true);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket()))));
+
+  SendPackets(/*timestamp=*/3, /*input=*/3, /*delay_next_output=*/true);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket()))));
+
+  SendPackets(/*timestamp=*/5, /*input=*/5, /*delay_next_output=*/false);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())),
+                  PairPacket(Timestamp(2), Pair(IntPacket(2), EmptyPacket())),
+                  PairPacket(Timestamp(3), Pair(IntPacket(3), EmptyPacket())),
+                  PairPacket(Timestamp(5), Pair(IntPacket(5), EmptyPacket()))));
+
+  SendPackets(/*timestamp=*/15, /*input=*/15, /*delay_next_output=*/false);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(
+          PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())),
+          PairPacket(Timestamp(2), Pair(IntPacket(2), EmptyPacket())),
+          PairPacket(Timestamp(3), Pair(IntPacket(3), EmptyPacket())),
+          PairPacket(Timestamp(5), Pair(IntPacket(5), EmptyPacket())),
+          PairPacket(Timestamp(15), Pair(IntPacket(15), IntPacket(5)))));
+
+  MP_EXPECT_OK(graph_.CloseAllInputStreams());
+  MP_EXPECT_OK(graph_.WaitUntilDone());
+}
+
+TEST_F(PreviousLoopbackCalculatorDelayBehaviorTest,
+       NonDelayedOutputFollowedByMultipleDelayedOutputs) {
+  SendPackets(/*timestamp=*/1, /*input=*/1, /*delay_next_output=*/false);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket()))));
+
+  SendPackets(/*timestamp=*/2, /*input=*/2, /*delay_next_output=*/true);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())),
+                  PairPacket(Timestamp(2), Pair(IntPacket(2), IntPacket(1)))));
+
+  SendPackets(/*timestamp=*/3, /*input=*/3, /*delay_next_output=*/true);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())),
+                  PairPacket(Timestamp(2), Pair(IntPacket(2), IntPacket(1)))));
+
+  SendPackets(/*timestamp=*/5, /*input=*/5, /*delay_next_output=*/false);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())),
+                  PairPacket(Timestamp(2), Pair(IntPacket(2), IntPacket(1))),
+                  PairPacket(Timestamp(3), Pair(IntPacket(3), EmptyPacket())),
+                  PairPacket(Timestamp(5), Pair(IntPacket(5), EmptyPacket()))));
+
+  SendPackets(/*timestamp=*/15, /*input=*/15, /*delay_next_output=*/false);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(
+          PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())),
+          PairPacket(Timestamp(2), Pair(IntPacket(2), IntPacket(1))),
+          PairPacket(Timestamp(3), Pair(IntPacket(3), EmptyPacket())),
+          PairPacket(Timestamp(5), Pair(IntPacket(5), EmptyPacket())),
+          PairPacket(Timestamp(15), Pair(IntPacket(15), IntPacket(5)))));
+
+  MP_EXPECT_OK(graph_.CloseAllInputStreams());
+  MP_EXPECT_OK(graph_.WaitUntilDone());
+}
+
 }  // anonymous namespace
 }  // namespace mediapipe

mediapipe/calculators/core/side_packet_to_stream_calculator.cc

@@ -28,55 +28,133 @@ using mediapipe::PacketTypeSet;
 using mediapipe::Timestamp;
 
 namespace {
+
+constexpr char kTagAtPreStream[] = "AT_PRESTREAM";
+constexpr char kTagAtPostStream[] = "AT_POSTSTREAM";
+constexpr char kTagAtZero[] = "AT_ZERO";
+constexpr char kTagAtTick[] = "AT_TICK";
+constexpr char kTagTick[] = "TICK";
+
 static std::map<std::string, Timestamp>* kTimestampMap = []() {
   auto* res = new std::map<std::string, Timestamp>();
-  res->emplace("AT_PRESTREAM", Timestamp::PreStream());
-  res->emplace("AT_POSTSTREAM", Timestamp::PostStream());
-  res->emplace("AT_ZERO", Timestamp(0));
+  res->emplace(kTagAtPreStream, Timestamp::PreStream());
+  res->emplace(kTagAtPostStream, Timestamp::PostStream());
+  res->emplace(kTagAtZero, Timestamp(0));
+  res->emplace(kTagAtTick, Timestamp::Unset());
   return res;
 }();
 
+template <typename CC>
+std::string GetOutputTag(const CC& cc) {
+  // Single output tag only is required by contract.
+  return *cc.Outputs().GetTags().begin();
+}
+
 }  // namespace
 
-// Outputs the single input_side_packet at the timestamp specified in the
-// output_stream tag. Valid tags are AT_PRESTREAM, AT_POSTSTREAM and AT_ZERO.
+// Outputs side packet(s) in corresponding output stream(s) with a particular
+// timestamp, depending on the tag used to define output stream(s). (One tag can
+// be used only.)
+//
+// Valid tags are AT_PRESTREAM, AT_POSTSTREAM, AT_ZERO and AT_TICK and
+// corresponding timestamps are Timestamp::PreStream(), Timestamp::PostStream(),
+// Timestamp(0) and timestamp of a packet received in TICK input.
+//
+// Examples:
+// node {
+//   calculator: "SidePacketToStreamCalculator"
+//   input_side_packet: "side_packet"
+//   output_stream: "AT_PRESTREAM:packet"
+// }
+//
+// node {
+//   calculator: "SidePacketToStreamCalculator"
+//   input_stream: "TICK:tick"
+//   input_side_packet: "side_packet"
+//   output_stream: "AT_TICK:packet"
+// }
 class SidePacketToStreamCalculator : public CalculatorBase {
  public:
   SidePacketToStreamCalculator() = default;
   ~SidePacketToStreamCalculator() override = default;
 
   static ::mediapipe::Status GetContract(CalculatorContract* cc);
+  ::mediapipe::Status Open(CalculatorContext* cc) override;
   ::mediapipe::Status Process(CalculatorContext* cc) override;
   ::mediapipe::Status Close(CalculatorContext* cc) override;
+
+ private:
+  bool is_tick_processing_ = false;
+  std::string output_tag_;
 };
 REGISTER_CALCULATOR(SidePacketToStreamCalculator);
 
 ::mediapipe::Status SidePacketToStreamCalculator::GetContract(
     CalculatorContract* cc) {
-  cc->InputSidePackets().Index(0).SetAny();
+  const auto& tags = cc->Outputs().GetTags();
+  RET_CHECK(tags.size() == 1 && kTimestampMap->count(*tags.begin()) == 1)
+      << "Only one of AT_PRESTREAM, AT_POSTSTREAM, AT_ZERO and AT_TICK tags is "
+         "allowed and required to specify output stream(s).";
+  RET_CHECK(
+      (cc->Outputs().HasTag(kTagAtTick) && cc->Inputs().HasTag(kTagTick)) ||
+      (!cc->Outputs().HasTag(kTagAtTick) && !cc->Inputs().HasTag(kTagTick)))
+      << "Either both of TICK and AT_TICK should be used or none of them.";
+  const std::string output_tag = GetOutputTag(*cc);
+  const int num_entries = cc->Outputs().NumEntries(output_tag);
+  RET_CHECK_EQ(num_entries, cc->InputSidePackets().NumEntries())
+      << "Same number of input side packets and output streams is required.";
+  for (int i = 0; i < num_entries; ++i) {
+    cc->InputSidePackets().Index(i).SetAny();
+    cc->Outputs()
+        .Get(output_tag, i)
+        .SetSameAs(cc->InputSidePackets().Index(i).GetSameAs());
+  }
 
-  std::set<std::string> tags = cc->Outputs().GetTags();
-  RET_CHECK_EQ(tags.size(), 1);
+  if (cc->Inputs().HasTag(kTagTick)) {
+    cc->Inputs().Tag(kTagTick).SetAny();
+  }
 
-  RET_CHECK_EQ(kTimestampMap->count(*tags.begin()), 1);
-  cc->Outputs().Tag(*tags.begin()).SetAny();
+  return ::mediapipe::OkStatus();
+}
 
+::mediapipe::Status SidePacketToStreamCalculator::Open(CalculatorContext* cc) {
+  output_tag_ = GetOutputTag(*cc);
+  if (cc->Inputs().HasTag(kTagTick)) {
+    is_tick_processing_ = true;
+    // Set offset, so output timestamp bounds are updated in response to TICK
+    // timestamp bound update.
+    cc->SetOffset(TimestampDiff(0));
+  }
   return ::mediapipe::OkStatus();
 }
 
 ::mediapipe::Status SidePacketToStreamCalculator::Process(
     CalculatorContext* cc) {
-  return mediapipe::tool::StatusStop();
+  if (is_tick_processing_) {
+    // TICK input is guaranteed to be non-empty, as it's the only input stream
+    // for this calculator.
+    const auto& timestamp = cc->Inputs().Tag(kTagTick).Value().Timestamp();
+    for (int i = 0; i < cc->Outputs().NumEntries(output_tag_); ++i) {
+      cc->Outputs()
+          .Get(output_tag_, i)
+          .AddPacket(cc->InputSidePackets().Index(i).At(timestamp));
+    }
+
+    return ::mediapipe::OkStatus();
+  }
+
+  return ::mediapipe::tool::StatusStop();
 }
 
 ::mediapipe::Status SidePacketToStreamCalculator::Close(CalculatorContext* cc) {
-  std::set<std::string> tags = cc->Outputs().GetTags();
-  RET_CHECK_EQ(tags.size(), 1);
-  const std::string& tag = *tags.begin();
-  RET_CHECK_EQ(kTimestampMap->count(tag), 1);
-  cc->Outputs().Tag(tag).AddPacket(
-      cc->InputSidePackets().Index(0).At(kTimestampMap->at(tag)));
-
+  if (!cc->Outputs().HasTag(kTagAtTick)) {
+    const auto& timestamp = kTimestampMap->at(output_tag_);
+    for (int i = 0; i < cc->Outputs().NumEntries(output_tag_); ++i) {
+      cc->Outputs()
+          .Get(output_tag_, i)
+          .AddPacket(cc->InputSidePackets().Index(i).At(timestamp));
+    }
+  }
   return ::mediapipe::OkStatus();
 }
 

mediapipe/calculators/core/side_packet_to_stream_calculator_test.cc

@@ -0,0 +1,275 @@
+// Copyright 2020 The MediaPipe Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <string>
+#include <vector>
+
+#include "absl/memory/memory.h"
+#include "absl/strings/match.h"
+#include "absl/strings/str_replace.h"
+#include "absl/strings/string_view.h"
+#include "mediapipe/framework/calculator_framework.h"
+#include "mediapipe/framework/port/gtest.h"
+#include "mediapipe/framework/port/integral_types.h"
+#include "mediapipe/framework/port/parse_text_proto.h"
+#include "mediapipe/framework/port/status.h"
+#include "mediapipe/framework/port/status_matchers.h"
+#include "mediapipe/framework/tool/options_util.h"
+
+namespace mediapipe {
+namespace {
+
+TEST(SidePacketToStreamCalculator, WrongConfig_MissingTick) {
+  CalculatorGraphConfig graph_config =
+      ParseTextProtoOrDie<CalculatorGraphConfig>(
+          R"(
+            input_stream: "tick"
+            input_side_packet: "side_packet"
+            output_stream: "packet"
+            node {
+              calculator: "SidePacketToStreamCalculator"
+              input_side_packet: "side_packet"
+              output_stream: "AT_TICK:packet"
+            }
+          )");
+  CalculatorGraph graph;
+  auto status = graph.Initialize(graph_config);
+  EXPECT_FALSE(status.ok());
+  EXPECT_PRED2(
+      absl::StrContains, status.message(),
+      "Either both of TICK and AT_TICK should be used or none of them.");
+}
+
+TEST(SidePacketToStreamCalculator, WrongConfig_NonExistentTag) {
+  CalculatorGraphConfig graph_config =
+      ParseTextProtoOrDie<CalculatorGraphConfig>(
+          R"(
+            input_stream: "tick"
+            input_side_packet: "side_packet"
+            output_stream: "packet"
+            node {
+              calculator: "SidePacketToStreamCalculator"
+              input_side_packet: "side_packet"
+              output_stream: "DOES_NOT_EXIST:packet"
+            }
+          )");
+  CalculatorGraph graph;
+  auto status = graph.Initialize(graph_config);
+  EXPECT_FALSE(status.ok());
+  EXPECT_PRED2(absl::StrContains, status.message(),
+               "Only one of AT_PRESTREAM, AT_POSTSTREAM, AT_ZERO and AT_TICK "
+               "tags is allowed and required to specify output stream(s).");
+}
+
+TEST(SidePacketToStreamCalculator, WrongConfig_MixedTags) {
+  CalculatorGraphConfig graph_config =
+      ParseTextProtoOrDie<CalculatorGraphConfig>(
+          R"(
+            input_stream: "tick"
+            input_side_packet: "side_packet0"
+            input_side_packet: "side_packet1"
+            node {
+              calculator: "SidePacketToStreamCalculator"
+              input_side_packet: "side_packet0"
+              input_side_packet: "side_packet1"
+              output_stream: "AT_TICK:packet0"
+              output_stream: "AT_PRE_STREAM:packet1"
+            }
+          )");
+  CalculatorGraph graph;
+  auto status = graph.Initialize(graph_config);
+  EXPECT_FALSE(status.ok());
+  EXPECT_PRED2(absl::StrContains, status.message(),
+               "Only one of AT_PRESTREAM, AT_POSTSTREAM, AT_ZERO and AT_TICK "
+               "tags is allowed and required to specify output stream(s).");
+}
+
+TEST(SidePacketToStreamCalculator, WrongConfig_NotEnoughSidePackets) {
+  CalculatorGraphConfig graph_config =
+      ParseTextProtoOrDie<CalculatorGraphConfig>(
+          R"(
+            input_side_packet: "side_packet0"
+            input_side_packet: "side_packet1"
+            node {
+              calculator: "SidePacketToStreamCalculator"
+              input_side_packet: "side_packet0"
+              output_stream: "AT_PRESTREAM:0:packet0"
+              output_stream: "AT_PRESTREAM:1:packet1"
+            }
+          )");
+  CalculatorGraph graph;
+  auto status = graph.Initialize(graph_config);
+  EXPECT_FALSE(status.ok());
+  EXPECT_PRED2(
+      absl::StrContains, status.message(),
+      "Same number of input side packets and output streams is required.");
+}
+
+TEST(SidePacketToStreamCalculator, WrongConfig_NotEnoughOutputStreams) {
+  CalculatorGraphConfig graph_config =
+      ParseTextProtoOrDie<CalculatorGraphConfig>(
+          R"(
+            input_side_packet: "side_packet0"
+            input_side_packet: "side_packet1"
+            node {
+              calculator: "SidePacketToStreamCalculator"
+              input_side_packet: "side_packet0"
+              input_side_packet: "side_packet1"
+              output_stream: "AT_PRESTREAM:packet0"
+            }
+          )");
+  CalculatorGraph graph;
+  auto status = graph.Initialize(graph_config);
+  EXPECT_FALSE(status.ok());
+  EXPECT_PRED2(
+      absl::StrContains, status.message(),
+      "Same number of input side packets and output streams is required.");
+}
+
+void DoTestNonAtTickOutputTag(absl::string_view tag,
+                              Timestamp expected_timestamp) {
+  CalculatorGraphConfig graph_config =
+      ParseTextProtoOrDie<CalculatorGraphConfig>(absl::StrReplaceAll(
+          R"(
+            input_side_packet: "side_packet"
+            output_stream: "packet"
+            node {
+              calculator: "SidePacketToStreamCalculator"
+              input_side_packet: "side_packet"
+              output_stream: "$tag:packet"
+            }
+          )",
+          {{"$tag", tag}}));
+  CalculatorGraph graph;
+  MP_ASSERT_OK(graph.Initialize(graph_config));
+  const int expected_value = 10;
+  std::vector<Packet> output_packets;
+  MP_ASSERT_OK(graph.ObserveOutputStream(
+      "packet", [&output_packets](const Packet& packet) {
+        output_packets.push_back(packet);
+        return ::mediapipe::OkStatus();
+      }));
+  MP_ASSERT_OK(
+      graph.StartRun({{"side_packet", MakePacket<int>(expected_value)}}));
+  MP_ASSERT_OK(graph.WaitForObservedOutput());
+
+  ASSERT_FALSE(output_packets.empty());
+  EXPECT_EQ(expected_timestamp, output_packets.back().Timestamp());
+  EXPECT_EQ(expected_value, output_packets.back().Get<int>());
+}
+
+TEST(SidePacketToStreamCalculator, NoAtTickOutputTags) {
+  DoTestNonAtTickOutputTag("AT_PRESTREAM", Timestamp::PreStream());
+  DoTestNonAtTickOutputTag("AT_POSTSTREAM", Timestamp::PostStream());
+  DoTestNonAtTickOutputTag("AT_ZERO", Timestamp(0));
+}
+
+TEST(SidePacketToStreamCalculator, AtTick) {
+  CalculatorGraphConfig graph_config =
+      ParseTextProtoOrDie<CalculatorGraphConfig>(
+          R"(
+            input_stream: "tick"
+            input_side_packet: "side_packet"
+            output_stream: "packet"
+            node {
+              calculator: "SidePacketToStreamCalculator"
+              input_stream: "TICK:tick"
+              input_side_packet: "side_packet"
+              output_stream: "AT_TICK:packet"
+            }
+          )");
+  std::vector<Packet> output_packets;
+  tool::AddVectorSink("packet", &graph_config, &output_packets);
+  CalculatorGraph graph;
+
+  MP_ASSERT_OK(graph.Initialize(graph_config));
+  const int expected_value = 20;
+  MP_ASSERT_OK(
+      graph.StartRun({{"side_packet", MakePacket<int>(expected_value)}}));
+
+  auto tick_and_verify = [&graph, &output_packets,
+                          expected_value](int at_timestamp) {
+    MP_ASSERT_OK(graph.AddPacketToInputStream(
+        "tick",
+        MakePacket<int>(/*doesn't matter*/ 1).At(Timestamp(at_timestamp))));
+    MP_ASSERT_OK(graph.WaitUntilIdle());
+
+    ASSERT_FALSE(output_packets.empty());
+    EXPECT_EQ(Timestamp(at_timestamp), output_packets.back().Timestamp());
+    EXPECT_EQ(expected_value, output_packets.back().Get<int>());
+  };
+
+  tick_and_verify(/*at_timestamp=*/0);
+  tick_and_verify(/*at_timestamp=*/1);
+  tick_and_verify(/*at_timestamp=*/128);
+  tick_and_verify(/*at_timestamp=*/1024);
+  tick_and_verify(/*at_timestamp=*/1025);
+}
+
+TEST(SidePacketToStreamCalculator, AtTick_MultipleSidePackets) {
+  CalculatorGraphConfig graph_config =
+      ParseTextProtoOrDie<CalculatorGraphConfig>(
+          R"(
+            input_stream: "tick"
+            input_side_packet: "side_packet0"
+            input_side_packet: "side_packet1"
+            output_stream: "packet0"
+            output_stream: "packet1"
+            node {
+              calculator: "SidePacketToStreamCalculator"
+              input_stream: "TICK:tick"
+              input_side_packet: "side_packet0"
+              input_side_packet: "side_packet1"
+              output_stream: "AT_TICK:0:packet0"
+              output_stream: "AT_TICK:1:packet1"
+            }
+          )");
+  std::vector<Packet> output_packets0;
+  tool::AddVectorSink("packet0", &graph_config, &output_packets0);
+  std::vector<Packet> output_packets1;
+  tool::AddVectorSink("packet1", &graph_config, &output_packets1);
+  CalculatorGraph graph;
+
+  MP_ASSERT_OK(graph.Initialize(graph_config));
+  const int expected_value0 = 20;
+  const int expected_value1 = 128;
+  MP_ASSERT_OK(
+      graph.StartRun({{"side_packet0", MakePacket<int>(expected_value0)},
+                      {"side_packet1", MakePacket<int>(expected_value1)}}));
+
+  auto tick_and_verify = [&graph, &output_packets0, &output_packets1,
+                          expected_value0, expected_value1](int at_timestamp) {
+    MP_ASSERT_OK(graph.AddPacketToInputStream(
+        "tick",
+        MakePacket<int>(/*doesn't matter*/ 1).At(Timestamp(at_timestamp))));
+    MP_ASSERT_OK(graph.WaitUntilIdle());
+
+    ASSERT_FALSE(output_packets0.empty());
+    ASSERT_FALSE(output_packets1.empty());
+
+    EXPECT_EQ(Timestamp(at_timestamp), output_packets0.back().Timestamp());
+    EXPECT_EQ(expected_value0, output_packets0.back().Get<int>());
+    EXPECT_EQ(Timestamp(at_timestamp), output_packets1.back().Timestamp());
+    EXPECT_EQ(expected_value1, output_packets1.back().Get<int>());
+  };
+
+  tick_and_verify(/*at_timestamp=*/0);
+  tick_and_verify(/*at_timestamp=*/1);
+  tick_and_verify(/*at_timestamp=*/128);
+  tick_and_verify(/*at_timestamp=*/1024);
+  tick_and_verify(/*at_timestamp=*/1025);
+}
+
+}  // namespace
+}  // namespace mediapipe

mediapipe/calculators/core/split_vector_calculator.cc

@@ -16,7 +16,9 @@
 
 #include <vector>
 
+#include "mediapipe/framework/formats/detection.pb.h"
 #include "mediapipe/framework/formats/landmark.pb.h"
+#include "mediapipe/framework/formats/matrix.h"
 #include "mediapipe/framework/formats/rect.pb.h"
 #include "tensorflow/lite/interpreter.h"
 
@@ -48,13 +50,25 @@ typedef SplitVectorCalculator<::mediapipe::NormalizedLandmark, false>
     SplitLandmarkVectorCalculator;
 REGISTER_CALCULATOR(SplitLandmarkVectorCalculator);
 
+typedef SplitVectorCalculator<::mediapipe::NormalizedLandmarkList, false>
+    SplitNormalizedLandmarkListVectorCalculator;
+REGISTER_CALCULATOR(SplitNormalizedLandmarkListVectorCalculator);
+
 typedef SplitVectorCalculator<::mediapipe::NormalizedRect, false>
     SplitNormalizedRectVectorCalculator;
 REGISTER_CALCULATOR(SplitNormalizedRectVectorCalculator);
 
+typedef SplitVectorCalculator<Matrix, false> SplitMatrixVectorCalculator;
+REGISTER_CALCULATOR(SplitMatrixVectorCalculator);
+
 #if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
 typedef SplitVectorCalculator<::tflite::gpu::gl::GlBuffer, true>
     MovableSplitGlBufferVectorCalculator;
 REGISTER_CALCULATOR(MovableSplitGlBufferVectorCalculator);
 #endif
+
+typedef SplitVectorCalculator<::mediapipe::Detection, false>
+    SplitDetectionVectorCalculator;
+REGISTER_CALCULATOR(SplitDetectionVectorCalculator);
+
 }  // namespace mediapipe

mediapipe/calculators/image/BUILD

@@ -86,6 +86,14 @@ mediapipe_cc_proto_library(
     deps = [":opencv_image_encoder_calculator_proto"],
 )
 
+mediapipe_cc_proto_library(
+    name = "opencv_encoded_image_to_image_frame_calculator_cc_proto",
+    srcs = ["opencv_encoded_image_to_image_frame_calculator.proto"],
+    cc_deps = ["//mediapipe/framework:calculator_cc_proto"],
+    visibility = ["//visibility:public"],
+    deps = [":opencv_encoded_image_to_image_frame_calculator_proto"],
+)
+
 mediapipe_cc_proto_library(
     name = "mask_overlay_calculator_cc_proto",
     srcs = ["mask_overlay_calculator.proto"],
@@ -172,6 +180,7 @@ cc_library(
     srcs = ["opencv_encoded_image_to_image_frame_calculator.cc"],
     visibility = ["//visibility:public"],
     deps = [
+        ":opencv_encoded_image_to_image_frame_calculator_cc_proto",
         "//mediapipe/framework:calculator_framework",
         "//mediapipe/framework/formats:image_frame_opencv",
         "//mediapipe/framework/port:opencv_imgcodecs",
@@ -346,9 +355,7 @@ cc_library(
         ],
         "//conditions:default": [],
     }),
-    visibility = [
-        "//visibility:public",
-    ],
+    visibility = ["//visibility:public"],
     deps = [
         ":image_cropping_calculator_cc_proto",
         "//mediapipe/framework:calculator_framework",
@@ -424,9 +431,12 @@ cc_library(
         ":recolor_calculator_cc_proto",
         "//mediapipe/framework:calculator_framework",
         "//mediapipe/framework/formats:image_frame",
+        "//mediapipe/framework/formats:image_frame_opencv",
         "//mediapipe/framework/port:status",
         "//mediapipe/framework/port:ret_check",
         "//mediapipe/util:color_cc_proto",
+        "//mediapipe/framework/port:opencv_core",
+        "//mediapipe/framework/port:opencv_imgproc",
     ] + select({
         "//mediapipe/gpu:disable_gpu": [],
         "//conditions:default": [
@@ -556,6 +566,27 @@ proto_library(
     deps = ["//mediapipe/framework:calculator_proto"],
 )
 
+proto_library(
+    name = "opencv_encoded_image_to_image_frame_calculator_proto",
+    srcs = ["opencv_encoded_image_to_image_frame_calculator.proto"],
+    visibility = ["//visibility:public"],
+    deps = ["//mediapipe/framework:calculator_proto"],
+)
+
+proto_library(
+    name = "feature_detector_calculator_proto",
+    srcs = ["feature_detector_calculator.proto"],
+    deps = ["//mediapipe/framework:calculator_proto"],
+)
+
+mediapipe_cc_proto_library(
+    name = "feature_detector_calculator_cc_proto",
+    srcs = ["feature_detector_calculator.proto"],
+    cc_deps = ["//mediapipe/framework:calculator_cc_proto"],
+    visibility = ["//visibility:public"],
+    deps = [":feature_detector_calculator_proto"],
+)
+
 cc_library(
     name = "mask_overlay_calculator",
     srcs = ["mask_overlay_calculator.cc"],
@@ -571,3 +602,30 @@ cc_library(
     ],
     alwayslink = 1,
 )
+
+cc_library(
+    name = "feature_detector_calculator",
+    srcs = ["feature_detector_calculator.cc"],
+    visibility = ["//mediapipe:__subpackages__"],
+    deps = [
+        ":feature_detector_calculator_cc_proto",
+        "//mediapipe/framework:calculator_framework",
+        "//mediapipe/framework/formats:image_frame",
+        "//mediapipe/framework/formats:image_frame_opencv",
+        "//mediapipe/framework/formats:landmark_cc_proto",
+        "//mediapipe/framework/formats:video_stream_header",
+        "//mediapipe/framework/port:integral_types",
+        "//mediapipe/framework/port:logging",
+        "//mediapipe/framework/port:opencv_core",
+        "//mediapipe/framework/port:opencv_features2d",
+        "//mediapipe/framework/port:opencv_imgproc",
+        "//mediapipe/framework/port:ret_check",
+        "//mediapipe/framework/port:status",
+        "//mediapipe/framework/port:threadpool",
+        "//mediapipe/framework/tool:options_util",
+        "@com_google_absl//absl/memory",
+        "@com_google_absl//absl/synchronization",
+        "@org_tensorflow//tensorflow/lite:framework",
+    ],
+    alwayslink = 1,
+)

mediapipe/calculators/image/color_convert_calculator.cc

@@ -75,6 +75,11 @@ class ColorConvertCalculator : public CalculatorBase {
   static ::mediapipe::Status GetContract(CalculatorContract* cc);
   ::mediapipe::Status Process(CalculatorContext* cc) override;
 
+  ::mediapipe::Status Open(CalculatorContext* cc) override {
+    cc->SetOffset(TimestampDiff(0));
+    return ::mediapipe::OkStatus();
+  }
+
  private:
   // Wrangles the appropriate inputs and outputs to perform the color
   // conversion. The ImageFrame on input_tag is converted using the

mediapipe/calculators/image/feature_detector_calculator.cc

@@ -0,0 +1,210 @@
+// Copyright 2020 The MediaPipe Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <memory>
+#include <vector>
+
+#include "absl/memory/memory.h"
+#include "absl/synchronization/blocking_counter.h"
+#include "mediapipe/calculators/image/feature_detector_calculator.pb.h"
+#include "mediapipe/framework/calculator_framework.h"
+#include "mediapipe/framework/formats/image_frame.h"
+#include "mediapipe/framework/formats/image_frame_opencv.h"
+#include "mediapipe/framework/formats/landmark.pb.h"
+#include "mediapipe/framework/formats/video_stream_header.h"
+#include "mediapipe/framework/port/integral_types.h"
+#include "mediapipe/framework/port/logging.h"
+#include "mediapipe/framework/port/opencv_core_inc.h"
+#include "mediapipe/framework/port/opencv_features2d_inc.h"
+#include "mediapipe/framework/port/opencv_imgproc_inc.h"
+#include "mediapipe/framework/port/ret_check.h"
+#include "mediapipe/framework/port/status.h"
+#include "mediapipe/framework/port/threadpool.h"
+#include "mediapipe/framework/tool/options_util.h"
+#include "tensorflow/lite/interpreter.h"
+
+namespace mediapipe {
+
+const char kOptionsTag[] = "OPTIONS";
+const int kPatchSize = 32;
+const int kNumThreads = 16;
+
+// A calculator to apply local feature detection.
+// Input stream:
+//   IMAGE: Input image frame of type ImageFrame from video stream.
+// Output streams:
+//   FEATURES: The detected keypoints from input image as vector<cv::KeyPoint>.
+//   PATCHES:  Optional output the extracted patches as vector<cv::Mat>
+class FeatureDetectorCalculator : public CalculatorBase {
+ public:
+  ~FeatureDetectorCalculator() override = default;
+
+  static ::mediapipe::Status GetContract(CalculatorContract* cc);
+
+  ::mediapipe::Status Open(CalculatorContext* cc) override;
+  ::mediapipe::Status Process(CalculatorContext* cc) override;
+
+ private:
+  FeatureDetectorCalculatorOptions options_;
+  cv::Ptr<cv::Feature2D> feature_detector_;
+  std::unique_ptr<::mediapipe::ThreadPool> pool_;
+
+  // Create image pyramid based on input image.
+  void ComputeImagePyramid(const cv::Mat& input_image,
+                           std::vector<cv::Mat>* image_pyramid);
+
+  // Extract the patch for single feature with image pyramid.
+  cv::Mat ExtractPatch(const cv::KeyPoint& feature,
+                       const std::vector<cv::Mat>& image_pyramid);
+};
+
+REGISTER_CALCULATOR(FeatureDetectorCalculator);
+
+::mediapipe::Status FeatureDetectorCalculator::GetContract(
+    CalculatorContract* cc) {
+  if (cc->Inputs().HasTag("IMAGE")) {
+    cc->Inputs().Tag("IMAGE").Set<ImageFrame>();
+  }
+  if (cc->Outputs().HasTag("FEATURES")) {
+    cc->Outputs().Tag("FEATURES").Set<std::vector<cv::KeyPoint>>();
+  }
+  if (cc->Outputs().HasTag("LANDMARKS")) {
+    cc->Outputs().Tag("LANDMARKS").Set<NormalizedLandmarkList>();
+  }
+  if (cc->Outputs().HasTag("PATCHES")) {
+    cc->Outputs().Tag("PATCHES").Set<std::vector<TfLiteTensor>>();
+  }
+  return ::mediapipe::OkStatus();
+}
+
+::mediapipe::Status FeatureDetectorCalculator::Open(CalculatorContext* cc) {
+  options_ =
+      tool::RetrieveOptions(cc->Options(), cc->InputSidePackets(), kOptionsTag)
+          .GetExtension(FeatureDetectorCalculatorOptions::ext);
+  feature_detector_ = cv::ORB::create(
+      options_.max_features(), options_.scale_factor(),
+      options_.pyramid_level(), kPatchSize - 1, 0, 2, cv::ORB::FAST_SCORE);
+  pool_ = absl::make_unique<::mediapipe::ThreadPool>("ThreadPool", kNumThreads);
+  pool_->StartWorkers();
+  return ::mediapipe::OkStatus();
+}
+
+::mediapipe::Status FeatureDetectorCalculator::Process(CalculatorContext* cc) {
+  const Timestamp& timestamp = cc->InputTimestamp();
+  if (timestamp == Timestamp::PreStream()) {
+    // Indicator packet.
+    return ::mediapipe::OkStatus();
+  }
+  InputStream* input_frame = &(cc->Inputs().Tag("IMAGE"));
+  cv::Mat input_view = formats::MatView(&input_frame->Get<ImageFrame>());
+  cv::Mat grayscale_view;
+  cv::cvtColor(input_view, grayscale_view, cv::COLOR_RGB2GRAY);
+
+  std::vector<cv::KeyPoint> keypoints;
+  feature_detector_->detect(grayscale_view, keypoints);
+  if (keypoints.size() > options_.max_features()) {
+    keypoints.resize(options_.max_features());
+  }
+
+  if (cc->Outputs().HasTag("FEATURES")) {
+    auto features_ptr = absl::make_unique<std::vector<cv::KeyPoint>>(keypoints);
+    cc->Outputs().Tag("FEATURES").Add(features_ptr.release(), timestamp);
+  }
+
+  if (cc->Outputs().HasTag("LANDMARKS")) {
+    auto landmarks_ptr = absl::make_unique<NormalizedLandmarkList>();
+    for (int j = 0; j < keypoints.size(); ++j) {
+      auto feature_landmark = landmarks_ptr->add_landmark();
+      feature_landmark->set_x(keypoints[j].pt.x / grayscale_view.cols);
+      feature_landmark->set_y(keypoints[j].pt.y / grayscale_view.rows);
+    }
+    cc->Outputs().Tag("LANDMARKS").Add(landmarks_ptr.release(), timestamp);
+  }
+
+  if (cc->Outputs().HasTag("PATCHES")) {
+    std::vector<cv::Mat> image_pyramid;
+    ComputeImagePyramid(grayscale_view, &image_pyramid);
+    std::vector<cv::Mat> patch_mat;
+    patch_mat.resize(keypoints.size());
+    absl::BlockingCounter counter(keypoints.size());
+    for (int i = 0; i < keypoints.size(); i++) {
+      pool_->Schedule(
+          [this, &image_pyramid, &keypoints, &patch_mat, i, &counter] {
+            patch_mat[i] = ExtractPatch(keypoints[i], image_pyramid);
+            counter.DecrementCount();
+          });
+    }
+    counter.Wait();
+    const int batch_size = options_.max_features();
+    auto patches = absl::make_unique<std::vector<TfLiteTensor>>();
+    TfLiteTensor tensor;
+    tensor.type = kTfLiteFloat32;
+    tensor.dims = TfLiteIntArrayCreate(4);
+    tensor.dims->data[0] = batch_size;
+    tensor.dims->data[1] = kPatchSize;
+    tensor.dims->data[2] = kPatchSize;
+    tensor.dims->data[3] = 1;
+    int num_bytes = batch_size * kPatchSize * kPatchSize * sizeof(float);
+    tensor.data.data = malloc(num_bytes);
+    tensor.bytes = num_bytes;
+    tensor.allocation_type = kTfLiteArenaRw;
+    float* tensor_buffer = tensor.data.f;
+    for (int i = 0; i < keypoints.size(); i++) {
+      for (int j = 0; j < patch_mat[i].rows; ++j) {
+        for (int k = 0; k < patch_mat[i].cols; ++k) {
+          *tensor_buffer++ = patch_mat[i].at<uchar>(j, k) / 128.0f - 1.0f;
+        }
+      }
+    }
+    for (int i = keypoints.size() * kPatchSize * kPatchSize; i < num_bytes / 4;
+         i++) {
+      *tensor_buffer++ = 0;
+    }
+
+    patches->emplace_back(tensor);
+    cc->Outputs().Tag("PATCHES").Add(patches.release(), timestamp);
+  }
+
+  return ::mediapipe::OkStatus();
+}
+
+void FeatureDetectorCalculator::ComputeImagePyramid(
+    const cv::Mat& input_image, std::vector<cv::Mat>* image_pyramid) {
+  cv::Mat tmp_image = input_image;
+  cv::Mat src_image = input_image;
+  for (int i = 0; i < options_.pyramid_level(); ++i) {
+    image_pyramid->push_back(src_image);
+    cv::resize(src_image, tmp_image, cv::Size(), 1.0f / options_.scale_factor(),
+               1.0f / options_.scale_factor());
+    src_image = tmp_image;
+  }
+}
+
+cv::Mat FeatureDetectorCalculator::ExtractPatch(
+    const cv::KeyPoint& feature, const std::vector<cv::Mat>& image_pyramid) {
+  cv::Mat img = image_pyramid[feature.octave];
+  float scale_factor = 1 / pow(options_.scale_factor(), feature.octave);
+  cv::Point2f center =
+      cv::Point2f(feature.pt.x * scale_factor, feature.pt.y * scale_factor);
+  cv::Mat rot = cv::getRotationMatrix2D(center, feature.angle, 1.0);
+  rot.at<double>(0, 2) += kPatchSize / 2 - center.x;
+  rot.at<double>(1, 2) += kPatchSize / 2 - center.y;
+  cv::Mat cropped_img;
+  // perform the affine transformation
+  cv::warpAffine(img, cropped_img, rot, cv::Size(kPatchSize, kPatchSize),
+                 cv::INTER_LINEAR);
+  return cropped_img;
+}
+
+}  // namespace mediapipe

mediapipe/calculators/image/feature_detector_calculator.proto

@@ -0,0 +1,24 @@
+// Options for FeatureDetectorCalculator
+syntax = "proto2";
+
+package mediapipe;
+
+import "mediapipe/framework/calculator.proto";
+
+message FeatureDetectorCalculatorOptions {
+  extend CalculatorOptions {
+    optional FeatureDetectorCalculatorOptions ext = 278741680;
+  }
+
+  // Set to true if output patches, otherwise only output cv::KeyPoint
+  optional bool output_patch = 1;
+
+  // The max number of detected features.
+  optional int32 max_features = 2 [default = 200];
+
+  // The number of pyramid levels.
+  optional int32 pyramid_level = 3 [default = 4];
+
+  // Pyramid decimation ratio.
+  optional float scale_factor = 4 [default = 1.2];
+}

mediapipe/calculators/image/image_cropping_calculator.cc

@@ -75,11 +75,28 @@ REGISTER_CALCULATOR(ImageCroppingCalculator);
   }
 #endif  //  !MEDIAPIPE_DISABLE_GPU
 
-  RET_CHECK(cc->Inputs().HasTag(kRectTag) ^ cc->Inputs().HasTag(kNormRectTag) ^
-            (cc->Options<mediapipe::ImageCroppingCalculatorOptions>()
-                 .has_norm_width() &&
-             cc->Options<mediapipe::ImageCroppingCalculatorOptions>()
-                 .has_norm_height()));
+  int flags = 0;
+  if (cc->Inputs().HasTag(kRectTag)) {
+    ++flags;
+  }
+  if (cc->Inputs().HasTag(kWidthTag) && cc->Inputs().HasTag(kHeightTag)) {
+    ++flags;
+  }
+  if (cc->Inputs().HasTag(kNormRectTag)) {
+    ++flags;
+  }
+  if (cc->Options<mediapipe::ImageCroppingCalculatorOptions>()
+          .has_norm_width() &&
+      cc->Options<mediapipe::ImageCroppingCalculatorOptions>()
+          .has_norm_height()) {
+    ++flags;
+  }
+  if (cc->Options<mediapipe::ImageCroppingCalculatorOptions>().has_width() &&
+      cc->Options<mediapipe::ImageCroppingCalculatorOptions>().has_height()) {
+    ++flags;
+  }
+  RET_CHECK(flags == 1) << "Illegal combination of input streams/options.";
+
   if (cc->Inputs().HasTag(kRectTag)) {
     cc->Inputs().Tag(kRectTag).Set<Rect>();
   }
@@ -119,6 +136,13 @@ REGISTER_CALCULATOR(ImageCroppingCalculator);
 #endif  //  !MEDIAPIPE_DISABLE_GPU
   }
 
+  // Validate border mode.
+  if (use_gpu_) {
+    MP_RETURN_IF_ERROR(ValidateBorderModeForGPU(cc));
+  } else {
+    MP_RETURN_IF_ERROR(ValidateBorderModeForCPU(cc));
+  }
+
   return ::mediapipe::OkStatus();
 }
 
@@ -162,6 +186,32 @@ REGISTER_CALCULATOR(ImageCroppingCalculator);
   return ::mediapipe::OkStatus();
 }
 
+::mediapipe::Status ImageCroppingCalculator::ValidateBorderModeForCPU(
+    CalculatorContext* cc) {
+  int border_mode;
+  return GetBorderModeForOpenCV(cc, &border_mode);
+}
+
+::mediapipe::Status ImageCroppingCalculator::ValidateBorderModeForGPU(
+    CalculatorContext* cc) {
+  mediapipe::ImageCroppingCalculatorOptions options =
+      cc->Options<mediapipe::ImageCroppingCalculatorOptions>();
+
+  switch (options.border_mode()) {
+    case mediapipe::ImageCroppingCalculatorOptions::BORDER_ZERO:
+      LOG(WARNING) << "BORDER_ZERO mode is not supported by GPU "
+                   << "implementation and will fall back into BORDER_REPLICATE";
+      break;
+    case mediapipe::ImageCroppingCalculatorOptions::BORDER_REPLICATE:
+      break;
+    default:
+      RET_CHECK_FAIL() << "Unsupported border mode for GPU: "
+                       << options.border_mode();
+  }
+
+  return ::mediapipe::OkStatus();
+}
+
 ::mediapipe::Status ImageCroppingCalculator::RenderCpu(CalculatorContext* cc) {
   if (cc->Inputs().Tag(kImageTag).IsEmpty()) {
     return ::mediapipe::OkStatus();
@@ -169,8 +219,14 @@ REGISTER_CALCULATOR(ImageCroppingCalculator);
   const auto& input_img = cc->Inputs().Tag(kImageTag).Get<ImageFrame>();
   cv::Mat input_mat = formats::MatView(&input_img);
 
-  auto [target_width, target_height, rect_center_x, rect_center_y, rotation] =
-      GetCropSpecs(cc, input_img.Width(), input_img.Height());
+  RectSpec specs = GetCropSpecs(cc, input_img.Width(), input_img.Height());
+  int target_width = specs.width, target_height = specs.height,
+      rect_center_x = specs.center_x, rect_center_y = specs.center_y;
+  float rotation = specs.rotation;
+
+  // Get border mode and value for OpenCV.
+  int border_mode;
+  MP_RETURN_IF_ERROR(GetBorderModeForOpenCV(cc, &border_mode));
 
   const cv::RotatedRect min_rect(cv::Point2f(rect_center_x, rect_center_y),
                                  cv::Size2f(target_width, target_height),
@@ -191,7 +247,9 @@ REGISTER_CALCULATOR(ImageCroppingCalculator);
       cv::getPerspectiveTransform(src_points, dst_points);
   cv::Mat cropped_image;
   cv::warpPerspective(input_mat, cropped_image, projection_matrix,
-                      cv::Size(min_rect.size.width, min_rect.size.height));
+                      cv::Size(min_rect.size.width, min_rect.size.height),
+                      /* flags = */ 0,
+                      /* borderMode = */ border_mode);
 
   std::unique_ptr<ImageFrame> output_frame(new ImageFrame(
       input_img.Format(), cropped_image.cols, cropped_image.rows));
@@ -347,8 +405,10 @@ void ImageCroppingCalculator::GetOutputDimensions(CalculatorContext* cc,
                                                   int src_width, int src_height,
                                                   int* dst_width,
                                                   int* dst_height) {
-  auto [crop_width, crop_height, x_center, y_center, rotation] =
-      GetCropSpecs(cc, src_width, src_height);
+  RectSpec specs = GetCropSpecs(cc, src_width, src_height);
+  int crop_width = specs.width, crop_height = specs.height,
+      x_center = specs.center_x, y_center = specs.center_y;
+  float rotation = specs.rotation;
 
   const float half_width = crop_width / 2.0f;
   const float half_height = crop_height / 2.0f;
@@ -453,6 +513,7 @@ RectSpec ImageCroppingCalculator::GetCropSpecs(const CalculatorContext* cc,
       rotation = options.rotation();
     }
   }
+
   return {
       .width = crop_width,
       .height = crop_height,
@@ -462,4 +523,24 @@ RectSpec ImageCroppingCalculator::GetCropSpecs(const CalculatorContext* cc,
   };
 }
 
+::mediapipe::Status ImageCroppingCalculator::GetBorderModeForOpenCV(
+    CalculatorContext* cc, int* border_mode) {
+  mediapipe::ImageCroppingCalculatorOptions options =
+      cc->Options<mediapipe::ImageCroppingCalculatorOptions>();
+
+  switch (options.border_mode()) {
+    case mediapipe::ImageCroppingCalculatorOptions::BORDER_ZERO:
+      *border_mode = cv::BORDER_CONSTANT;
+      break;
+    case mediapipe::ImageCroppingCalculatorOptions::BORDER_REPLICATE:
+      *border_mode = cv::BORDER_REPLICATE;
+      break;
+    default:
+      RET_CHECK_FAIL() << "Unsupported border mode for CPU: "
+                       << options.border_mode();
+  }
+
+  return ::mediapipe::OkStatus();
+}
+
 }  // namespace mediapipe

mediapipe/calculators/image/image_cropping_calculator.h

@@ -36,6 +36,7 @@
 // Note: input_stream values take precedence over options defined in the graph.
 //
 namespace mediapipe {
+
 struct RectSpec {
   int width;
   int height;
@@ -63,12 +64,16 @@ class ImageCroppingCalculator : public CalculatorBase {
                                int src_height);
 
  private:
+  ::mediapipe::Status ValidateBorderModeForCPU(CalculatorContext* cc);
+  ::mediapipe::Status ValidateBorderModeForGPU(CalculatorContext* cc);
   ::mediapipe::Status RenderCpu(CalculatorContext* cc);
   ::mediapipe::Status RenderGpu(CalculatorContext* cc);
   ::mediapipe::Status InitGpu(CalculatorContext* cc);
   void GlRender();
   void GetOutputDimensions(CalculatorContext* cc, int src_width, int src_height,
                            int* dst_width, int* dst_height);
+  ::mediapipe::Status GetBorderModeForOpenCV(CalculatorContext* cc,
+                                             int* border_mode);
 
   mediapipe::ImageCroppingCalculatorOptions options_;
 

mediapipe/calculators/image/image_cropping_calculator.proto

@@ -40,4 +40,15 @@ message ImageCroppingCalculatorOptions {
   // The (0, 0) point is at the (top, left) corner.
   optional float norm_center_x = 6 [default = 0];
   optional float norm_center_y = 7 [default = 0];
+
+  enum BorderMode {
+    // First unspecified value is required by the guideline. See details here:
+    // https://developers.google.com/protocol-buffers/docs/style#enums
+    BORDER_UNSPECIFIED = 0;
+    BORDER_ZERO = 1;
+    BORDER_REPLICATE = 2;
+  }
+
+  // Specifies behaviour for crops that go beyond image borders.
+  optional BorderMode border_mode = 8 [default = BORDER_ZERO];
 }

mediapipe/calculators/image/image_cropping_calculator_test.cc

@@ -56,11 +56,11 @@ TEST(ImageCroppingCalculatorTest, GetCroppingDimensionsNormal) {
             }
           )");
 
-  auto calculator_state =
-      CalculatorState("Node", 0, "Calculator", calculator_node, nullptr);
-  auto cc =
-      CalculatorContext(&calculator_state, tool::CreateTagMap({}).ValueOrDie(),
-                        tool::CreateTagMap({}).ValueOrDie());
+  auto calculator_state = absl::make_unique<CalculatorState>(
+      "Node", 0, "Calculator", calculator_node, nullptr);
+  auto cc = absl::make_unique<CalculatorContext>(
+      calculator_state.get(), tool::CreateTagMap({}).ValueOrDie(),
+      tool::CreateTagMap({}).ValueOrDie());
 
   RectSpec expectRect = {
       .width = 60,
@@ -69,9 +69,9 @@ TEST(ImageCroppingCalculatorTest, GetCroppingDimensionsNormal) {
       .center_y = 50,
       .rotation = 0.3,
   };
-  EXPECT_EQ(
-      ImageCroppingCalculator::GetCropSpecs(&cc, input_width, input_height),
-      expectRect);
+  EXPECT_EQ(ImageCroppingCalculator::GetCropSpecs(cc.get(), input_width,
+                                                  input_height),
+            expectRect);
 }  // TEST
 
 // Test when (width height) + (norm_width norm_height) are set in options.
@@ -96,11 +96,11 @@ TEST(ImageCroppingCalculatorTest, RedundantSpecInOptions) {
             }
           )");
 
-  auto calculator_state =
-      CalculatorState("Node", 0, "Calculator", calculator_node, nullptr);
-  auto cc =
-      CalculatorContext(&calculator_state, tool::CreateTagMap({}).ValueOrDie(),
-                        tool::CreateTagMap({}).ValueOrDie());
+  auto calculator_state = absl::make_unique<CalculatorState>(
+      "Node", 0, "Calculator", calculator_node, nullptr);
+  auto cc = absl::make_unique<CalculatorContext>(
+      calculator_state.get(), tool::CreateTagMap({}).ValueOrDie(),
+      tool::CreateTagMap({}).ValueOrDie());
   RectSpec expectRect = {
       .width = 50,
       .height = 50,
@@ -108,9 +108,9 @@ TEST(ImageCroppingCalculatorTest, RedundantSpecInOptions) {
       .center_y = 50,
       .rotation = 0.3,
   };
-  EXPECT_EQ(
-      ImageCroppingCalculator::GetCropSpecs(&cc, input_width, input_height),
-      expectRect);
+  EXPECT_EQ(ImageCroppingCalculator::GetCropSpecs(cc.get(), input_width,
+                                                  input_height),
+            expectRect);
 }  // TEST
 
 // Test when WIDTH HEIGHT are set from input stream,
@@ -138,16 +138,16 @@ TEST(ImageCroppingCalculatorTest, RedundantSpectWithInputStream) {
             }
           )");
 
-  auto calculator_state =
-      CalculatorState("Node", 0, "Calculator", calculator_node, nullptr);
+  auto calculator_state = absl::make_unique<CalculatorState>(
+      "Node", 0, "Calculator", calculator_node, nullptr);
   auto inputTags = tool::CreateTagMap({
                                           "HEIGHT:0:crop_height",
                                           "WIDTH:0:crop_width",
                                       })
                        .ValueOrDie();
-  auto cc = CalculatorContext(&calculator_state, inputTags,
-                              tool::CreateTagMap({}).ValueOrDie());
-  auto& inputs = cc.Inputs();
+  auto cc = absl::make_unique<CalculatorContext>(
+      calculator_state.get(), inputTags, tool::CreateTagMap({}).ValueOrDie());
+  auto& inputs = cc->Inputs();
   inputs.Tag(kHeightTag).Value() = MakePacket<int>(1);
   inputs.Tag(kWidthTag).Value() = MakePacket<int>(1);
   RectSpec expectRect = {
@@ -157,9 +157,9 @@ TEST(ImageCroppingCalculatorTest, RedundantSpectWithInputStream) {
       .center_y = 50,
       .rotation = 0.3,
   };
-  EXPECT_EQ(
-      ImageCroppingCalculator::GetCropSpecs(&cc, input_width, input_height),
-      expectRect);
+  EXPECT_EQ(ImageCroppingCalculator::GetCropSpecs(cc.get(), input_width,
+                                                  input_height),
+            expectRect);
 }  // TEST
 
 // Test when RECT is set from input stream,
@@ -186,15 +186,15 @@ TEST(ImageCroppingCalculatorTest, RedundantSpecWithInputStream) {
             }
           )");
 
-  auto calculator_state =
-      CalculatorState("Node", 0, "Calculator", calculator_node, nullptr);
+  auto calculator_state = absl::make_unique<CalculatorState>(
+      "Node", 0, "Calculator", calculator_node, nullptr);
   auto inputTags = tool::CreateTagMap({
                                           "RECT:0:rect",
                                       })
                        .ValueOrDie();
-  auto cc = CalculatorContext(&calculator_state, inputTags,
-                              tool::CreateTagMap({}).ValueOrDie());
-  auto& inputs = cc.Inputs();
+  auto cc = absl::make_unique<CalculatorContext>(
+      calculator_state.get(), inputTags, tool::CreateTagMap({}).ValueOrDie());
+  auto& inputs = cc->Inputs();
   mediapipe::Rect rect = ParseTextProtoOrDie<mediapipe::Rect>(
       R"(
         width: 1 height: 1 x_center: 40 y_center: 40 rotation: 0.5
@@ -207,9 +207,9 @@ TEST(ImageCroppingCalculatorTest, RedundantSpecWithInputStream) {
       .center_y = 40,
       .rotation = 0.5,
   };
-  EXPECT_EQ(
-      ImageCroppingCalculator::GetCropSpecs(&cc, input_width, input_height),
-      expectRect);
+  EXPECT_EQ(ImageCroppingCalculator::GetCropSpecs(cc.get(), input_width,
+                                                  input_height),
+            expectRect);
 }  // TEST
 
 }  // namespace

mediapipe/calculators/image/image_transformation_calculator.cc

@@ -104,6 +104,14 @@ mediapipe::ScaleMode_Mode ParseScaleMode(
 //   to be a multiple of 90 degrees. If provided, it overrides the
 //   ROTATION_DEGREES input side packet.
 //
+//   FLIP_HORIZONTALLY (optional): Whether to flip image horizontally or not. If
+//   provided, it overrides the FLIP_HORIZONTALLY input side packet and/or
+//   corresponding field in the calculator options.
+//
+//   FLIP_VERTICALLY (optional): Whether to flip image vertically or not. If
+//   provided, it overrides the FLIP_VERTICALLY input side packet and/or
+//   corresponding field in the calculator options.
+//
 // Output:
 //   One of the following two tags:
 //   IMAGE - ImageFrame representing the output image.
@@ -129,6 +137,12 @@ mediapipe::ScaleMode_Mode ParseScaleMode(
 //   degrees. It has to be a multiple of 90 degrees. It overrides the
 //   corresponding field in the calculator options.
 //
+//   FLIP_HORIZONTALLY (optional): Whether to flip image horizontally or not.
+//   It overrides the corresponding field in the calculator options.
+//
+//   FLIP_VERTICALLY (optional): Whether to flip image vertically or not.
+//   It overrides the corresponding field in the calculator options.
+//
 // Calculator options (see image_transformation_calculator.proto):
 //   output_width, output_height - (optional) Desired scaled image size.
 //   rotation_mode - (optional) Rotation in multiples of 90 degrees.
@@ -167,6 +181,8 @@ class ImageTransformationCalculator : public CalculatorBase {
   int output_height_ = 0;
   mediapipe::RotationMode_Mode rotation_;
   mediapipe::ScaleMode_Mode scale_mode_;
+  bool flip_horizontally_ = false;
+  bool flip_vertically_ = false;
 
   bool use_gpu_ = false;
 #if !defined(MEDIAPIPE_DISABLE_GPU)
@@ -203,6 +219,12 @@ REGISTER_CALCULATOR(ImageTransformationCalculator);
   if (cc->Inputs().HasTag("ROTATION_DEGREES")) {
     cc->Inputs().Tag("ROTATION_DEGREES").Set<int>();
   }
+  if (cc->Inputs().HasTag("FLIP_HORIZONTALLY")) {
+    cc->Inputs().Tag("FLIP_HORIZONTALLY").Set<bool>();
+  }
+  if (cc->Inputs().HasTag("FLIP_VERTICALLY")) {
+    cc->Inputs().Tag("FLIP_VERTICALLY").Set<bool>();
+  }
 
   if (cc->InputSidePackets().HasTag("OUTPUT_DIMENSIONS")) {
     cc->InputSidePackets().Tag("OUTPUT_DIMENSIONS").Set<DimensionsPacketType>();
@@ -210,6 +232,12 @@ REGISTER_CALCULATOR(ImageTransformationCalculator);
   if (cc->InputSidePackets().HasTag("ROTATION_DEGREES")) {
     cc->InputSidePackets().Tag("ROTATION_DEGREES").Set<int>();
   }
+  if (cc->InputSidePackets().HasTag("FLIP_HORIZONTALLY")) {
+    cc->InputSidePackets().Tag("FLIP_HORIZONTALLY").Set<bool>();
+  }
+  if (cc->InputSidePackets().HasTag("FLIP_VERTICALLY")) {
+    cc->InputSidePackets().Tag("FLIP_VERTICALLY").Set<bool>();
+  }
 
   if (cc->Outputs().HasTag("LETTERBOX_PADDING")) {
     cc->Outputs().Tag("LETTERBOX_PADDING").Set<std::array<float, 4>>();
@@ -245,6 +273,7 @@ REGISTER_CALCULATOR(ImageTransformationCalculator);
     output_width_ = options_.output_width();
     output_height_ = options_.output_height();
   }
+
   if (cc->InputSidePackets().HasTag("ROTATION_DEGREES")) {
     rotation_ = DegreesToRotationMode(
         cc->InputSidePackets().Tag("ROTATION_DEGREES").Get<int>());
@@ -252,6 +281,20 @@ REGISTER_CALCULATOR(ImageTransformationCalculator);
     rotation_ = options_.rotation_mode();
   }
 
+  if (cc->InputSidePackets().HasTag("FLIP_HORIZONTALLY")) {
+    flip_horizontally_ =
+        cc->InputSidePackets().Tag("FLIP_HORIZONTALLY").Get<bool>();
+  } else {
+    flip_horizontally_ = options_.flip_horizontally();
+  }
+
+  if (cc->InputSidePackets().HasTag("FLIP_VERTICALLY")) {
+    flip_vertically_ =
+        cc->InputSidePackets().Tag("FLIP_VERTICALLY").Get<bool>();
+  } else {
+    flip_vertically_ = options_.flip_vertically();
+  }
+
   scale_mode_ = ParseScaleMode(options_.scale_mode(), DEFAULT_SCALE_MODE);
 
   if (use_gpu_) {
@@ -268,12 +311,37 @@ REGISTER_CALCULATOR(ImageTransformationCalculator);
 
 ::mediapipe::Status ImageTransformationCalculator::Process(
     CalculatorContext* cc) {
+  // Override values if specified so.
+  if (cc->Inputs().HasTag("ROTATION_DEGREES") &&
+      !cc->Inputs().Tag("ROTATION_DEGREES").IsEmpty()) {
+    rotation_ =
+        DegreesToRotationMode(cc->Inputs().Tag("ROTATION_DEGREES").Get<int>());
+  }
+  if (cc->Inputs().HasTag("FLIP_HORIZONTALLY") &&
+      !cc->Inputs().Tag("FLIP_HORIZONTALLY").IsEmpty()) {
+    flip_horizontally_ = cc->Inputs().Tag("FLIP_HORIZONTALLY").Get<bool>();
+  }
+  if (cc->Inputs().HasTag("FLIP_VERTICALLY") &&
+      !cc->Inputs().Tag("FLIP_VERTICALLY").IsEmpty()) {
+    flip_vertically_ = cc->Inputs().Tag("FLIP_VERTICALLY").Get<bool>();
+  }
+
   if (use_gpu_) {
 #if !defined(MEDIAPIPE_DISABLE_GPU)
+    if (cc->Inputs().Tag("IMAGE_GPU").IsEmpty()) {
+      // Image is missing, hence no way to produce output image. (Timestamp
+      // bound will be updated automatically.)
+      return ::mediapipe::OkStatus();
+    }
     return helper_.RunInGlContext(
         [this, cc]() -> ::mediapipe::Status { return RenderGpu(cc); });
 #endif  //  !MEDIAPIPE_DISABLE_GPU
   } else {
+    if (cc->Inputs().Tag("IMAGE").IsEmpty()) {
+      // Image is missing, hence no way to produce output image. (Timestamp
+      // bound will be updated automatically.)
+      return ::mediapipe::OkStatus();
+    }
     return RenderCpu(cc);
   }
   return ::mediapipe::OkStatus();
@@ -360,11 +428,6 @@ REGISTER_CALCULATOR(ImageTransformationCalculator);
         .Add(padding.release(), cc->InputTimestamp());
   }
 
-  if (cc->InputSidePackets().HasTag("ROTATION_DEGREES")) {
-    rotation_ = DegreesToRotationMode(
-        cc->InputSidePackets().Tag("ROTATION_DEGREES").Get<int>());
-  }
-
   cv::Mat rotated_mat;
   const int angle = RotationModeToDegrees(rotation_);
   cv::Point2f src_center(scaled_mat.cols / 2.0, scaled_mat.rows / 2.0);
@@ -372,11 +435,9 @@ REGISTER_CALCULATOR(ImageTransformationCalculator);
   cv::warpAffine(scaled_mat, rotated_mat, rotation_mat, scaled_mat.size());
 
   cv::Mat flipped_mat;
-  if (options_.flip_horizontally() || options_.flip_vertically()) {
+  if (flip_horizontally_ || flip_vertically_) {
     const int flip_code =
-        options_.flip_horizontally() && options_.flip_vertically()
-            ? -1
-            : options_.flip_horizontally();
+        flip_horizontally_ && flip_vertically_ ? -1 : flip_horizontally_;
     cv::flip(rotated_mat, flipped_mat, flip_code);
   } else {
     flipped_mat = rotated_mat;
@@ -450,11 +511,6 @@ REGISTER_CALCULATOR(ImageTransformationCalculator);
   }
   RET_CHECK(renderer) << "Unsupported input texture type";
 
-  if (cc->InputSidePackets().HasTag("ROTATION_DEGREES")) {
-    rotation_ = DegreesToRotationMode(
-        cc->InputSidePackets().Tag("ROTATION_DEGREES").Get<int>());
-  }
-
   mediapipe::FrameScaleMode scale_mode = mediapipe::FrameScaleModeFromProto(
       scale_mode_, mediapipe::FrameScaleMode::kStretch);
   mediapipe::FrameRotation rotation =
@@ -469,7 +525,7 @@ REGISTER_CALCULATOR(ImageTransformationCalculator);
 
   MP_RETURN_IF_ERROR(renderer->GlRender(
       src1.width(), src1.height(), dst.width(), dst.height(), scale_mode,
-      rotation, options_.flip_horizontally(), options_.flip_vertically(),
+      rotation, flip_horizontally_, flip_vertically_,
       /*flip_texture=*/false));
 
   glActiveTexture(GL_TEXTURE1);

mediapipe/calculators/image/opencv_encoded_image_to_image_frame_calculator.cc

@@ -12,6 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+#include "mediapipe/calculators/image/opencv_encoded_image_to_image_frame_calculator.pb.h"
 #include "mediapipe/framework/calculator_framework.h"
 #include "mediapipe/framework/formats/image_frame_opencv.h"
 #include "mediapipe/framework/port/opencv_imgcodecs_inc.h"
@@ -34,7 +35,11 @@ namespace mediapipe {
 class OpenCvEncodedImageToImageFrameCalculator : public CalculatorBase {
  public:
   static ::mediapipe::Status GetContract(CalculatorContract* cc);
+  ::mediapipe::Status Open(CalculatorContext* cc) override;
   ::mediapipe::Status Process(CalculatorContext* cc) override;
+
+ private:
+  mediapipe::OpenCvEncodedImageToImageFrameCalculatorOptions options_;
 };
 
 ::mediapipe::Status OpenCvEncodedImageToImageFrameCalculator::GetContract(
@@ -44,13 +49,29 @@ class OpenCvEncodedImageToImageFrameCalculator : public CalculatorBase {
   return ::mediapipe::OkStatus();
 }
 
+::mediapipe::Status OpenCvEncodedImageToImageFrameCalculator::Open(
+    CalculatorContext* cc) {
+  options_ =
+      cc->Options<mediapipe::OpenCvEncodedImageToImageFrameCalculatorOptions>();
+  return ::mediapipe::OkStatus();
+}
+
 ::mediapipe::Status OpenCvEncodedImageToImageFrameCalculator::Process(
     CalculatorContext* cc) {
   const std::string& contents = cc->Inputs().Index(0).Get<std::string>();
   const std::vector<char> contents_vector(contents.begin(), contents.end());
-  cv::Mat decoded_mat =
-      cv::imdecode(contents_vector, -1 /* return the loaded image as-is */);
-
+  cv::Mat decoded_mat;
+  if (options_.apply_orientation_from_exif_data()) {
+    // We want to respect the orientation from the EXIF data, which
+    // IMREAD_UNCHANGED ignores, but otherwise we want to be as permissive as
+    // possible with our reading flags. Therefore, we use IMREAD_ANYCOLOR and
+    // IMREAD_ANYDEPTH.
+    decoded_mat = cv::imdecode(contents_vector,
+                               cv::IMREAD_ANYCOLOR | cv::IMREAD_ANYDEPTH);
+  } else {
+    // Return the loaded image as-is
+    decoded_mat = cv::imdecode(contents_vector, cv::IMREAD_UNCHANGED);
+  }
   ImageFormat::Format image_format = ImageFormat::UNKNOWN;
   cv::Mat output_mat;
   switch (decoded_mat.channels()) {
@@ -70,7 +91,8 @@ class OpenCvEncodedImageToImageFrameCalculator : public CalculatorBase {
              << "Unsupported number of channels: " << decoded_mat.channels();
   }
   std::unique_ptr<ImageFrame> output_frame = absl::make_unique<ImageFrame>(
-      image_format, decoded_mat.size().width, decoded_mat.size().height);
+      image_format, decoded_mat.size().width, decoded_mat.size().height,
+      ImageFrame::kGlDefaultAlignmentBoundary);
   output_mat.copyTo(formats::MatView(output_frame.get()));
   cc->Outputs().Index(0).Add(output_frame.release(), cc->InputTimestamp());
   return ::mediapipe::OkStatus();

mediapipe/calculators/image/opencv_encoded_image_to_image_frame_calculator.proto

@@ -0,0 +1,30 @@
+// Copyright 2020 The MediaPipe Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+syntax = "proto2";
+
+package mediapipe;
+
+import "mediapipe/framework/calculator.proto";
+
+message OpenCvEncodedImageToImageFrameCalculatorOptions {
+  extend CalculatorOptions {
+    optional OpenCvEncodedImageToImageFrameCalculatorOptions ext = 303447308;
+  }
+
+  // If set, we will attempt to automatically apply the orientation specified by
+  // the image's EXIF data when loading the image. Otherwise, the image data
+  // will be loaded as-is.
+  optional bool apply_orientation_from_exif_data = 1 [default = false];
+}

mediapipe/calculators/image/recolor_calculator.cc

@@ -17,6 +17,9 @@
 #include "mediapipe/calculators/image/recolor_calculator.pb.h"
 #include "mediapipe/framework/calculator_framework.h"
 #include "mediapipe/framework/formats/image_frame.h"
+#include "mediapipe/framework/formats/image_frame_opencv.h"
+#include "mediapipe/framework/port/opencv_core_inc.h"
+#include "mediapipe/framework/port/opencv_imgproc_inc.h"
 #include "mediapipe/framework/port/ret_check.h"
 #include "mediapipe/framework/port/status.h"
 #include "mediapipe/util/color.pb.h"
@@ -39,8 +42,6 @@ namespace mediapipe {
 // The luminance of the input image is used to adjust the blending weight,
 // to help preserve image textures.
 //
-// TODO implement cpu support.
-//
 // Inputs:
 //   One of the following IMAGE tags:
 //   IMAGE: An ImageFrame input image, RGB or RGBA.
@@ -71,6 +72,8 @@ namespace mediapipe {
 //    }
 //  }
 //
+// Note: Cannot mix-match CPU & GPU inputs/outputs.
+//       CPU-in & CPU-out <or> GPU-in & GPU-out
 class RecolorCalculator : public CalculatorBase {
  public:
   RecolorCalculator() = default;
@@ -138,6 +141,11 @@ REGISTER_CALCULATOR(RecolorCalculator);
     cc->Outputs().Tag("IMAGE").Set<ImageFrame>();
   }
 
+  // Confirm only one of the input streams is present.
+  RET_CHECK(cc->Inputs().HasTag("IMAGE") ^ cc->Inputs().HasTag("IMAGE_GPU"));
+  // Confirm only one of the output streams is present.
+  RET_CHECK(cc->Outputs().HasTag("IMAGE") ^ cc->Outputs().HasTag("IMAGE_GPU"));
+
   if (use_gpu) {
 #if !defined(MEDIAPIPE_DISABLE_GPU)
     MP_RETURN_IF_ERROR(mediapipe::GlCalculatorHelper::UpdateContract(cc));
@@ -193,7 +201,62 @@ REGISTER_CALCULATOR(RecolorCalculator);
 }
 
 ::mediapipe::Status RecolorCalculator::RenderCpu(CalculatorContext* cc) {
-  return ::mediapipe::UnimplementedError("CPU support is not implemented yet.");
+  if (cc->Inputs().Tag("MASK").IsEmpty()) {
+    return ::mediapipe::OkStatus();
+  }
+  // Get inputs and setup output.
+  const auto& input_img = cc->Inputs().Tag("IMAGE").Get<ImageFrame>();
+  const auto& mask_img = cc->Inputs().Tag("MASK").Get<ImageFrame>();
+
+  cv::Mat input_mat = formats::MatView(&input_img);
+  cv::Mat mask_mat = formats::MatView(&mask_img);
+
+  RET_CHECK(input_mat.channels() == 3);  // RGB only.
+
+  if (mask_mat.channels() > 1) {
+    std::vector<cv::Mat> channels;
+    cv::split(mask_mat, channels);
+    if (mask_channel_ == mediapipe::RecolorCalculatorOptions_MaskChannel_ALPHA)
+      mask_mat = channels[3];
+    else
+      mask_mat = channels[0];
+  }
+  cv::Mat mask_full;
+  cv::resize(mask_mat, mask_full, input_mat.size());
+
+  auto output_img = absl::make_unique<ImageFrame>(
+      input_img.Format(), input_mat.cols, input_mat.rows);
+  cv::Mat output_mat = mediapipe::formats::MatView(output_img.get());
+
+  // From GPU shader:
+  /*
+      vec4 weight = texture2D(mask, sample_coordinate);
+      vec4 color1 = texture2D(frame, sample_coordinate);
+      vec4 color2 = vec4(recolor, 1.0);
+
+      float luminance = dot(color1.rgb, vec3(0.299, 0.587, 0.114));
+      float mix_value = weight.MASK_COMPONENT * luminance;
+
+      fragColor = mix(color1, color2, mix_value);
+  */
+  for (int i = 0; i < output_mat.rows; ++i) {
+    for (int j = 0; j < output_mat.cols; ++j) {
+      float weight = mask_full.at<uchar>(i, j) * (1.0 / 255.0);
+      cv::Vec3f color1 = input_mat.at<cv::Vec3b>(i, j);
+      cv::Vec3f color2 = {color_[0], color_[1], color_[2]};
+
+      float luminance =
+          (color1[0] * 0.299 + color1[1] * 0.587 + color1[2] * 0.114) / 255;
+      float mix_value = weight * luminance;
+
+      cv::Vec3b mix_color = color1 * (1.0 - mix_value) + color2 * mix_value;
+      output_mat.at<cv::Vec3b>(i, j) = mix_color;
+    }
+  }
+
+  cc->Outputs().Tag("IMAGE").Add(output_img.release(), cc->InputTimestamp());
+
+  return ::mediapipe::OkStatus();
 }
 
 ::mediapipe::Status RecolorCalculator::RenderGpu(CalculatorContext* cc) {
@@ -303,9 +366,9 @@ void RecolorCalculator::GlRender() {
 
   if (!options.has_color()) RET_CHECK_FAIL() << "Missing color option.";
 
-  color_.push_back(options.color().r() / 255.0);
-  color_.push_back(options.color().g() / 255.0);
-  color_.push_back(options.color().b() / 255.0);
+  color_.push_back(options.color().r());
+  color_.push_back(options.color().g());
+  color_.push_back(options.color().b());
 
   return ::mediapipe::OkStatus();
 }
@@ -378,8 +441,8 @@ void RecolorCalculator::GlRender() {
   glUseProgram(program_);
   glUniform1i(glGetUniformLocation(program_, "frame"), 1);
   glUniform1i(glGetUniformLocation(program_, "mask"), 2);
-  glUniform3f(glGetUniformLocation(program_, "recolor"), color_[0], color_[1],
-              color_[2]);
+  glUniform3f(glGetUniformLocation(program_, "recolor"), color_[0] / 255.0,
+              color_[1] / 255.0, color_[2] / 255.0);
 #endif  //  !MEDIAPIPE_DISABLE_GPU
 
   return ::mediapipe::OkStatus();

mediapipe/calculators/image/scale_image_calculator.cc

@@ -260,11 +260,11 @@ ScaleImageCalculator::~ScaleImageCalculator() {}
                                       &crop_width_, &crop_height_,  //
                                       &col_start_, &row_start_));
   MP_RETURN_IF_ERROR(
-      scale_image::FindOutputDimensions(crop_width_, crop_height_,            //
-                                        options_.target_width(),              //
-                                        options_.target_height(),             //
-                                        options_.preserve_aspect_ratio(),     //
-                                        options_.scale_to_multiple_of_two(),  //
+      scale_image::FindOutputDimensions(crop_width_, crop_height_,         //
+                                        options_.target_width(),           //
+                                        options_.target_height(),          //
+                                        options_.preserve_aspect_ratio(),  //
+                                        options_.scale_to_multiple_of(),   //
                                         &output_width_, &output_height_));
   MP_RETURN_IF_ERROR(FindInterpolationAlgorithm(options_.algorithm(),
                                                 &interpolation_algorithm_));
@@ -361,17 +361,21 @@ ScaleImageCalculator::~ScaleImageCalculator() {}
       output_format_ = input_format_;
     }
 
+    const bool is_positive_and_even =
+        (options_.scale_to_multiple_of() >= 1) &&
+        (options_.scale_to_multiple_of() % 2 == 0);
+
     if (output_format_ == ImageFormat::YCBCR420P) {
-      RET_CHECK(options_.scale_to_multiple_of_two())
+      RET_CHECK(is_positive_and_even)
           << "ScaleImageCalculator always outputs width and height that are "
              "divisible by 2 when output format is YCbCr420P. To scale to "
              "width and height of odd numbers, the output format must be SRGB.";
     } else if (options_.preserve_aspect_ratio()) {
-      RET_CHECK(options_.scale_to_multiple_of_two())
+      RET_CHECK(options_.scale_to_multiple_of() == 2)
           << "ScaleImageCalculator always outputs width and height that are "
-             "divisible by 2 when perserving aspect ratio. To scale to width "
-             "and height of odd numbers, please set "
-             "preserve_aspect_ratio to false.";
+             "divisible by 2 when preserving aspect ratio. If you'd like to "
+             "set scale_to_multiple_of to something other than 2, please "
+             "set preserve_aspect_ratio to false.";
     }
 
     if (input_width_ > 0 && input_height_ > 0 &&

mediapipe/calculators/image/scale_image_calculator.proto

@@ -11,9 +11,10 @@ import "mediapipe/framework/formats/image_format.proto";
 // 2) Scale and convert the image to fit inside target_width x target_height
 //    using the specified scaling algorithm.  (maintaining the aspect
 //    ratio if preserve_aspect_ratio is true).
-// The output width and height will be divisible by 2. It is possible to output
-// width and height that are odd number when the output format is SRGB and not
-// perserving the aspect ratio. See scale_to_multiple_of_two option for details.
+// The output width and height will be divisible by 2, by default. It is
+// possible to output width and height that are odd numbers when the output
+// format is SRGB and the aspect ratio is left unpreserved. See
+// scale_to_multiple_of for details.
 message ScaleImageCalculatorOptions {
   extend CalculatorOptions {
     optional ScaleImageCalculatorOptions ext = 66237115;
@@ -23,7 +24,7 @@ message ScaleImageCalculatorOptions {
   // depending on the other options below.  If unset, use the same width
   // or height as the input.  If only one is set then determine the other
   // from the aspect ratio (after cropping).  The output width and height
-  // will be divisible by 2.
+  // will be divisible by 2, by default.
   optional int32 target_width = 1;
   optional int32 target_height = 2;
 
@@ -31,7 +32,8 @@ message ScaleImageCalculatorOptions {
   // fits inside the box represented by target_width and target_height.
   // Otherwise it is scaled to fit target_width and target_height
   // completely.  In any case, the aspect ratio that is preserved is
-  // that after cropping to the minimum/maximum aspect ratio.
+  // that after cropping to the minimum/maximum aspect ratio. Additionally, if
+  // true, the output width and height will be divisible by 2.
   optional bool preserve_aspect_ratio = 3 [default = true];
 
   // If ratio is positive, crop the image to this minimum and maximum
@@ -95,11 +97,13 @@ message ScaleImageCalculatorOptions {
   // SRGB or YCBCR420P.
   optional ImageFormat.Format input_format = 12;
 
-  // If true, the output width and height will be divisible by 2. Otherwise it
-  // will use the exact specified output width and height, which is only
-  // supported when the output format is SRGB and preserve_aspect_ratio option
-  // is set to false.
-  optional bool scale_to_multiple_of_two = 13 [default = true];
+  // If set to 2, the target width and height will be rounded-down
+  // to the nearest even number. If set to any positive value other than 2,
+  // preserve_aspect_ratio must be false and the target width and height will be
+  // rounded-down to multiples of the given value. If set to any value less than
+  // 1, it will be treated like 1.
+  // NOTE: If set to an odd number, the output format must be SRGB.
+  optional int32 scale_to_multiple_of = 13 [default = 2];
 
   // If true, assume the input YUV is BT.709 (this is the HDTV standard, so most
   // content is likely using it). If false use the previous assumption of BT.601

mediapipe/calculators/image/scale_image_utils.cc

@@ -88,17 +88,27 @@ double ParseRational(const std::string& rational) {
   return ::mediapipe::OkStatus();
 }
 
-::mediapipe::Status FindOutputDimensions(int input_width,                //
-                                         int input_height,               //
-                                         int target_width,               //
-                                         int target_height,              //
-                                         bool preserve_aspect_ratio,     //
-                                         bool scale_to_multiple_of_two,  //
+::mediapipe::Status FindOutputDimensions(int input_width,             //
+                                         int input_height,            //
+                                         int target_width,            //
+                                         int target_height,           //
+                                         bool preserve_aspect_ratio,  //
+                                         int scale_to_multiple_of,    //
                                          int* output_width,
                                          int* output_height) {
   CHECK(output_width);
   CHECK(output_height);
 
+  if (preserve_aspect_ratio) {
+    RET_CHECK(scale_to_multiple_of == 2)
+        << "FindOutputDimensions always outputs width and height that are "
+           "divisible by 2 when preserving aspect ratio. If you'd like to "
+           "set scale_to_multiple_of to something other than 2, please "
+           "set preserve_aspect_ratio to false.";
+  }
+
+  if (scale_to_multiple_of < 1) scale_to_multiple_of = 1;
+
   if (!preserve_aspect_ratio || (target_width <= 0 && target_height <= 0)) {
     if (target_width <= 0) {
       target_width = input_width;
@@ -106,13 +116,13 @@ double ParseRational(const std::string& rational) {
     if (target_height <= 0) {
       target_height = input_height;
     }
-    if (scale_to_multiple_of_two) {
-      *output_width = (target_width / 2) * 2;
-      *output_height = (target_height / 2) * 2;
-    } else {
-      *output_width = target_width;
-      *output_height = target_height;
-    }
+
+    target_width -= target_width % scale_to_multiple_of;
+    target_height -= target_height % scale_to_multiple_of;
+
+    *output_width = target_width;
+    *output_height = target_height;
+
     return ::mediapipe::OkStatus();
   }
 

mediapipe/calculators/image/scale_image_utils.h

@@ -35,17 +35,19 @@ namespace scale_image {
                                        int* col_start, int* row_start);
 
 // Given an input width and height, a target width and height, whether to
-// preserve the aspect ratio, and whether to round down to a multiple of 2,
-// determine the output width and height. If target_width or target_height is
-// non-positive, then they will be set to the input_width and input_height
-// respectively. The output_width and output_height will be reduced as necessary
-// to preserve_aspect_ratio and to scale_to_multipe_of_two if these options are
-// specified.
+// preserve the aspect ratio, and whether to round-down to the multiple of a
+// given number nearest to the targets, determine the output width and height.
+// If target_width or target_height is non-positive, then they will be set to
+// the input_width and input_height respectively. If scale_to_multiple_of is
+// less than 1, it will be treated like 1. The output_width and
+// output_height will be reduced as necessary to preserve_aspect_ratio if the
+// option is specified. If preserving the aspect ratio is desired, you must set
+// scale_to_multiple_of to 2.
 ::mediapipe::Status FindOutputDimensions(int input_width, int input_height,  //
                                          int target_width,
-                                         int target_height,              //
-                                         bool preserve_aspect_ratio,     //
-                                         bool scale_to_multiple_of_two,  //
+                                         int target_height,           //
+                                         bool preserve_aspect_ratio,  //
+                                         int scale_to_multiple_of,    //
                                          int* output_width, int* output_height);
 
 }  // namespace scale_image

mediapipe/calculators/image/scale_image_utils_test.cc

@@ -79,49 +79,49 @@ TEST(ScaleImageUtilsTest, FindOutputDimensionsPreserveRatio) {
   int output_width;
   int output_height;
   // Not scale.
-  MP_ASSERT_OK(FindOutputDimensions(200, 100, -1, -1, true, true, &output_width,
+  MP_ASSERT_OK(FindOutputDimensions(200, 100, -1, -1, true, 2, &output_width,
                                     &output_height));
   EXPECT_EQ(200, output_width);
   EXPECT_EQ(100, output_height);
   // Not scale with odd input size.
-  MP_ASSERT_OK(FindOutputDimensions(201, 101, -1, -1, false, false,
-                                    &output_width, &output_height));
+  MP_ASSERT_OK(FindOutputDimensions(201, 101, -1, -1, false, 1, &output_width,
+                                    &output_height));
   EXPECT_EQ(201, output_width);
   EXPECT_EQ(101, output_height);
   // Scale down by 1/2.
-  MP_ASSERT_OK(FindOutputDimensions(200, 100, 100, -1, true, true,
-                                    &output_width, &output_height));
+  MP_ASSERT_OK(FindOutputDimensions(200, 100, 100, -1, true, 2, &output_width,
+                                    &output_height));
   EXPECT_EQ(100, output_width);
   EXPECT_EQ(50, output_height);
   // Scale up, doubling dimensions.
-  MP_ASSERT_OK(FindOutputDimensions(200, 100, -1, 200, true, true,
-                                    &output_width, &output_height));
+  MP_ASSERT_OK(FindOutputDimensions(200, 100, -1, 200, true, 2, &output_width,
+                                    &output_height));
   EXPECT_EQ(400, output_width);
   EXPECT_EQ(200, output_height);
   // Fits a 2:1 image into a 150 x 150 box. Output dimensions are always
   // visible by 2.
-  MP_ASSERT_OK(FindOutputDimensions(200, 100, 150, 150, true, true,
-                                    &output_width, &output_height));
+  MP_ASSERT_OK(FindOutputDimensions(200, 100, 150, 150, true, 2, &output_width,
+                                    &output_height));
   EXPECT_EQ(150, output_width);
   EXPECT_EQ(74, output_height);
   // Fits a 2:1 image into a 400 x 50 box.
-  MP_ASSERT_OK(FindOutputDimensions(200, 100, 400, 50, true, true,
-                                    &output_width, &output_height));
+  MP_ASSERT_OK(FindOutputDimensions(200, 100, 400, 50, true, 2, &output_width,
+                                    &output_height));
   EXPECT_EQ(100, output_width);
   EXPECT_EQ(50, output_height);
   // Scale to multiple number with odd targe size.
-  MP_ASSERT_OK(FindOutputDimensions(200, 100, 101, -1, true, true,
-                                    &output_width, &output_height));
+  MP_ASSERT_OK(FindOutputDimensions(200, 100, 101, -1, true, 2, &output_width,
+                                    &output_height));
   EXPECT_EQ(100, output_width);
   EXPECT_EQ(50, output_height);
   // Scale to multiple number with odd targe size.
-  MP_ASSERT_OK(FindOutputDimensions(200, 100, 101, -1, true, false,
-                                    &output_width, &output_height));
+  MP_ASSERT_OK(FindOutputDimensions(200, 100, 101, -1, true, 2, &output_width,
+                                    &output_height));
   EXPECT_EQ(100, output_width);
   EXPECT_EQ(50, output_height);
   // Scale to odd size.
-  MP_ASSERT_OK(FindOutputDimensions(200, 100, 151, 101, false, false,
-                                    &output_width, &output_height));
+  MP_ASSERT_OK(FindOutputDimensions(200, 100, 151, 101, false, 1, &output_width,
+                                    &output_height));
   EXPECT_EQ(151, output_width);
   EXPECT_EQ(101, output_height);
 }
@@ -131,22 +131,62 @@ TEST(ScaleImageUtilsTest, FindOutputDimensionsNoAspectRatio) {
   int output_width;
   int output_height;
   // Scale width only.
-  MP_ASSERT_OK(FindOutputDimensions(200, 100, 100, -1, false, true,
-                                    &output_width, &output_height));
+  MP_ASSERT_OK(FindOutputDimensions(200, 100, 100, -1, false, 2, &output_width,
+                                    &output_height));
   EXPECT_EQ(100, output_width);
   EXPECT_EQ(100, output_height);
   // Scale height only.
-  MP_ASSERT_OK(FindOutputDimensions(200, 100, -1, 200, false, true,
-                                    &output_width, &output_height));
+  MP_ASSERT_OK(FindOutputDimensions(200, 100, -1, 200, false, 2, &output_width,
+                                    &output_height));
   EXPECT_EQ(200, output_width);
   EXPECT_EQ(200, output_height);
   // Scale both dimensions.
-  MP_ASSERT_OK(FindOutputDimensions(200, 100, 150, 200, false, true,
-                                    &output_width, &output_height));
+  MP_ASSERT_OK(FindOutputDimensions(200, 100, 150, 200, false, 2, &output_width,
+                                    &output_height));
   EXPECT_EQ(150, output_width);
   EXPECT_EQ(200, output_height);
 }
 
+// Tests scale_to_multiple_of.
+TEST(ScaleImageUtilsTest, FindOutputDimensionsDownScaleToMultipleOf) {
+  int output_width;
+  int output_height;
+  // Set no targets, downscale to a multiple of 8.
+  MP_ASSERT_OK(FindOutputDimensions(100, 100, -1, -1, false, 8, &output_width,
+                                    &output_height));
+  EXPECT_EQ(96, output_width);
+  EXPECT_EQ(96, output_height);
+  // Set width target, downscale to a multiple of 8.
+  MP_ASSERT_OK(FindOutputDimensions(200, 100, 100, -1, false, 8, &output_width,
+                                    &output_height));
+  EXPECT_EQ(96, output_width);
+  EXPECT_EQ(96, output_height);
+  // Set height target, downscale to a multiple of 8.
+  MP_ASSERT_OK(FindOutputDimensions(201, 101, -1, 201, false, 8, &output_width,
+                                    &output_height));
+  EXPECT_EQ(200, output_width);
+  EXPECT_EQ(200, output_height);
+  // Set both targets, downscale to a multiple of 8.
+  MP_ASSERT_OK(FindOutputDimensions(200, 100, 150, 200, false, 8, &output_width,
+                                    &output_height));
+  EXPECT_EQ(144, output_width);
+  EXPECT_EQ(200, output_height);
+  // Doesn't throw error if keep aspect is true and downscale multiple is 2.
+  MP_ASSERT_OK(FindOutputDimensions(200, 100, 400, 200, true, 2, &output_width,
+                                    &output_height));
+  EXPECT_EQ(400, output_width);
+  EXPECT_EQ(200, output_height);
+  // Throws error if keep aspect is true, but downscale multiple is not 2.
+  ASSERT_THAT(FindOutputDimensions(200, 100, 400, 200, true, 4, &output_width,
+                                   &output_height),
+              testing::Not(testing::status::IsOk()));
+  // Downscaling to multiple ignored if multiple is less than 2.
+  MP_ASSERT_OK(FindOutputDimensions(200, 100, 401, 201, false, 1, &output_width,
+                                    &output_height));
+  EXPECT_EQ(401, output_width);
+  EXPECT_EQ(201, output_height);
+}
+
 }  // namespace
 }  // namespace scale_image
 }  // namespace mediapipe

mediapipe/calculators/tensorflow/BUILD

@@ -138,7 +138,7 @@ mediapipe_cc_proto_library(
     srcs = ["image_frame_to_tensor_calculator.proto"],
     cc_deps = [
         "//mediapipe/framework:calculator_cc_proto",
-        "@org_tensorflow//tensorflow/core:protos_all",
+        "@org_tensorflow//tensorflow/core:protos_all_cc",
     ],
     visibility = ["//visibility:public"],
     deps = [":image_frame_to_tensor_calculator_proto"],
@@ -173,7 +173,7 @@ mediapipe_cc_proto_library(
     srcs = ["pack_media_sequence_calculator.proto"],
     cc_deps = [
         "//mediapipe/framework:calculator_cc_proto",
-        "@org_tensorflow//tensorflow/core:protos_all",
+        "@org_tensorflow//tensorflow/core:protos_all_cc",
     ],
     visibility = ["//visibility:public"],
     deps = [":pack_media_sequence_calculator_proto"],
@@ -192,7 +192,7 @@ mediapipe_cc_proto_library(
     srcs = ["tensorflow_session_from_frozen_graph_generator.proto"],
     cc_deps = [
         "//mediapipe/framework:packet_generator_cc_proto",
-        "@org_tensorflow//tensorflow/core:protos_all",
+        "@org_tensorflow//tensorflow/core:protos_all_cc",
     ],
     visibility = ["//visibility:public"],
     deps = [":tensorflow_session_from_frozen_graph_generator_proto"],
@@ -203,7 +203,7 @@ mediapipe_cc_proto_library(
     srcs = ["tensorflow_session_from_frozen_graph_calculator.proto"],
     cc_deps = [
         "//mediapipe/framework:calculator_cc_proto",
-        "@org_tensorflow//tensorflow/core:protos_all",
+        "@org_tensorflow//tensorflow/core:protos_all_cc",
     ],
     visibility = ["//visibility:public"],
     deps = [":tensorflow_session_from_frozen_graph_calculator_proto"],
@@ -277,7 +277,7 @@ mediapipe_cc_proto_library(
     srcs = ["vector_int_to_tensor_calculator_options.proto"],
     cc_deps = [
         "//mediapipe/framework:calculator_cc_proto",
-        "@org_tensorflow//tensorflow/core:protos_all",
+        "@org_tensorflow//tensorflow/core:protos_all_cc",
     ],
     visibility = ["//visibility:public"],
     deps = [":vector_int_to_tensor_calculator_options_proto"],
@@ -408,7 +408,7 @@ cc_library(
         "//mediapipe/util/sequence:media_sequence",
         "//mediapipe/util/sequence:media_sequence_util",
         "@com_google_absl//absl/strings",
-        "@org_tensorflow//tensorflow/core:protos_all",
+        "@org_tensorflow//tensorflow/core:protos_all_cc",
     ],
     alwayslink = 1,
 )
@@ -423,7 +423,7 @@ cc_library(
         "//mediapipe/framework:calculator_framework",
         "//mediapipe/framework/port:ret_check",
         "//mediapipe/framework/port:status",
-        "@org_tensorflow//tensorflow/core:protos_all",
+        "@org_tensorflow//tensorflow/core:protos_all_cc",
     ],
     alwayslink = 1,
 )
@@ -654,7 +654,7 @@ cc_library(
         "//mediapipe/framework/port:ret_check",
         "//mediapipe/framework/port:status",
         "@org_tensorflow//tensorflow/core:lib",
-        "@org_tensorflow//tensorflow/core:protos_all",
+        "@org_tensorflow//tensorflow/core:protos_all_cc",
     ],
     alwayslink = 1,
 )
@@ -695,7 +695,7 @@ cc_library(
         "//mediapipe/util:audio_decoder_cc_proto",
         "//mediapipe/util/sequence:media_sequence",
         "@com_google_absl//absl/strings",
-        "@org_tensorflow//tensorflow/core:protos_all",
+        "@org_tensorflow//tensorflow/core:protos_all_cc",
     ],
     alwayslink = 1,
 )
@@ -737,7 +737,7 @@ cc_library(
         "//mediapipe/framework:calculator_framework",
         "//mediapipe/framework:packet",
         "//mediapipe/framework/port:status",
-        "@org_tensorflow//tensorflow/core:protos_all",
+        "@org_tensorflow//tensorflow/core:protos_all_cc",
     ],
     alwayslink = 1,
 )
@@ -745,6 +745,7 @@ cc_library(
 cc_test(
     name = "graph_tensors_packet_generator_test",
     srcs = ["graph_tensors_packet_generator_test.cc"],
+    linkstatic = 1,
     deps = [
         ":graph_tensors_packet_generator",
         ":graph_tensors_packet_generator_cc_proto",
@@ -761,6 +762,7 @@ cc_test(
     name = "image_frame_to_tensor_calculator_test",
     size = "small",
     srcs = ["image_frame_to_tensor_calculator_test.cc"],
+    linkstatic = 1,
     deps = [
         ":image_frame_to_tensor_calculator",
         "//mediapipe/framework:calculator_framework",
@@ -777,6 +779,7 @@ cc_test(
     name = "matrix_to_tensor_calculator_test",
     size = "small",
     srcs = ["matrix_to_tensor_calculator_test.cc"],
+    linkstatic = 1,
     deps = [
         ":matrix_to_tensor_calculator",
         ":matrix_to_tensor_calculator_options_cc_proto",
@@ -793,6 +796,7 @@ cc_test(
     name = "lapped_tensor_buffer_calculator_test",
     size = "small",
     srcs = ["lapped_tensor_buffer_calculator_test.cc"],
+    linkstatic = 1,
     deps = [
         ":lapped_tensor_buffer_calculator",
         ":lapped_tensor_buffer_calculator_cc_proto",
@@ -801,7 +805,7 @@ cc_test(
         "//mediapipe/framework/port:gtest_main",
         "@com_google_absl//absl/memory",
         "@org_tensorflow//tensorflow/core:framework",
-        "@org_tensorflow//tensorflow/core:protos_all",
+        "@org_tensorflow//tensorflow/core:protos_all_cc",
     ],
 )
 
@@ -840,7 +844,7 @@ cc_test(
         "//mediapipe/util/sequence:media_sequence",
         "@com_google_absl//absl/memory",
         "@com_google_absl//absl/strings",
-        "@org_tensorflow//tensorflow/core:protos_all",
+        "@org_tensorflow//tensorflow/core:protos_all_cc",
     ],
 )
 
@@ -867,7 +871,7 @@ cc_test(
         "@com_google_absl//absl/strings",
         "@org_tensorflow//tensorflow/core:direct_session",
         "@org_tensorflow//tensorflow/core:framework",
-        "@org_tensorflow//tensorflow/core:protos_all",
+        "@org_tensorflow//tensorflow/core:protos_all_cc",
         "@org_tensorflow//tensorflow/core:testlib",
         "@org_tensorflow//tensorflow/core/kernels:conv_ops",
         "@org_tensorflow//tensorflow/core/kernels:math",
@@ -897,7 +901,7 @@ cc_test(
         "@com_google_absl//absl/strings",
         "@org_tensorflow//tensorflow/core:direct_session",
         "@org_tensorflow//tensorflow/core:framework",
-        "@org_tensorflow//tensorflow/core:protos_all",
+        "@org_tensorflow//tensorflow/core:protos_all_cc",
         "@org_tensorflow//tensorflow/core:testlib",
         "@org_tensorflow//tensorflow/core/kernels:conv_ops",
         "@org_tensorflow//tensorflow/core/kernels:math",
@@ -956,6 +960,7 @@ cc_test(
 cc_test(
     name = "tensor_squeeze_dimensions_calculator_test",
     srcs = ["tensor_squeeze_dimensions_calculator_test.cc"],
+    linkstatic = 1,
     deps = [
         ":tensor_squeeze_dimensions_calculator",
         ":tensor_squeeze_dimensions_calculator_cc_proto",
@@ -963,7 +968,7 @@ cc_test(
         "//mediapipe/framework:calculator_runner",
         "//mediapipe/framework/port:gtest_main",
         "@org_tensorflow//tensorflow/core:framework",
-        "@org_tensorflow//tensorflow/core:protos_all",
+        "@org_tensorflow//tensorflow/core:protos_all_cc",
     ],
 )
 
@@ -971,6 +976,7 @@ cc_test(
     name = "tensor_to_image_frame_calculator_test",
     size = "small",
     srcs = ["tensor_to_image_frame_calculator_test.cc"],
+    linkstatic = 1,
     deps = [
         ":tensor_to_image_frame_calculator",
         ":tensor_to_image_frame_calculator_cc_proto",
@@ -979,7 +985,7 @@ cc_test(
         "//mediapipe/framework/formats:image_frame",
         "//mediapipe/framework/port:gtest_main",
         "@org_tensorflow//tensorflow/core:framework",
-        "@org_tensorflow//tensorflow/core:protos_all",
+        "@org_tensorflow//tensorflow/core:protos_all_cc",
     ],
 )
 
@@ -987,6 +993,7 @@ cc_test(
     name = "tensor_to_matrix_calculator_test",
     size = "small",
     srcs = ["tensor_to_matrix_calculator_test.cc"],
+    linkstatic = 1,
     deps = [
         ":tensor_to_matrix_calculator",
         ":tensor_to_matrix_calculator_cc_proto",
@@ -996,13 +1003,14 @@ cc_test(
         "//mediapipe/framework/formats:time_series_header_cc_proto",
         "//mediapipe/framework/port:gtest_main",
         "@org_tensorflow//tensorflow/core:framework",
-        "@org_tensorflow//tensorflow/core:protos_all",
+        "@org_tensorflow//tensorflow/core:protos_all_cc",
     ],
 )
 
 cc_test(
     name = "tensor_to_vector_float_calculator_test",
     srcs = ["tensor_to_vector_float_calculator_test.cc"],
+    linkstatic = 1,
     deps = [
         ":tensor_to_vector_float_calculator",
         ":tensor_to_vector_float_calculator_options_cc_proto",
@@ -1010,7 +1018,7 @@ cc_test(
         "//mediapipe/framework:calculator_runner",
         "//mediapipe/framework/port:gtest_main",
         "@org_tensorflow//tensorflow/core:framework",
-        "@org_tensorflow//tensorflow/core:protos_all",
+        "@org_tensorflow//tensorflow/core:protos_all_cc",
     ],
 )
 
@@ -1030,13 +1038,14 @@ cc_test(
         "//mediapipe/util/sequence:media_sequence",
         "@com_google_absl//absl/memory",
         "@com_google_absl//absl/strings",
-        "@org_tensorflow//tensorflow/core:protos_all",
+        "@org_tensorflow//tensorflow/core:protos_all_cc",
     ],
 )
 
 cc_test(
     name = "vector_int_to_tensor_calculator_test",
     srcs = ["vector_int_to_tensor_calculator_test.cc"],
+    linkstatic = 1,
     deps = [
         ":vector_int_to_tensor_calculator",
         ":vector_int_to_tensor_calculator_options_cc_proto",
@@ -1044,13 +1053,14 @@ cc_test(
         "//mediapipe/framework:calculator_runner",
         "//mediapipe/framework/port:gtest_main",
         "@org_tensorflow//tensorflow/core:framework",
-        "@org_tensorflow//tensorflow/core:protos_all",
+        "@org_tensorflow//tensorflow/core:protos_all_cc",
     ],
 )
 
 cc_test(
     name = "vector_float_to_tensor_calculator_test",
     srcs = ["vector_float_to_tensor_calculator_test.cc"],
+    linkstatic = 1,
     deps = [
         ":vector_float_to_tensor_calculator",
         ":vector_float_to_tensor_calculator_options_cc_proto",
@@ -1058,7 +1068,7 @@ cc_test(
         "//mediapipe/framework:calculator_runner",
         "//mediapipe/framework/port:gtest_main",
         "@org_tensorflow//tensorflow/core:framework",
-        "@org_tensorflow//tensorflow/core:protos_all",
+        "@org_tensorflow//tensorflow/core:protos_all_cc",
     ],
 )
 
@@ -1100,6 +1110,7 @@ cc_test(
         ],
         "//mediapipe:android": [
             "@org_tensorflow//tensorflow/core:android_tensorflow_lib_with_ops_lite_proto_no_rtti_lib",
+            "@org_tensorflow//tensorflow/core:android_tensorflow_test_lib",
         ],
         "//mediapipe:ios": [
             "@org_tensorflow//tensorflow/core:ios_tensorflow_test_lib",

mediapipe/calculators/tensorflow/tensorflow_session_from_saved_model_calculator.cc

@@ -17,7 +17,7 @@
 #if !defined(__ANDROID__)
 #include "mediapipe/framework/port/file_helpers.h"
 #endif
-#include "absl/strings/substitute.h"
+#include "absl/strings/str_replace.h"
 #include "mediapipe/calculators/tensorflow/tensorflow_session.h"
 #include "mediapipe/calculators/tensorflow/tensorflow_session_from_saved_model_calculator.pb.h"
 #include "mediapipe/framework/calculator_framework.h"
@@ -51,10 +51,11 @@ static constexpr char kStringSavedModelPath[] = "STRING_SAVED_MODEL_PATH";
 #endif
 }
 
-// If options.convert_signature_to_tags() will convert letters to uppercase
-// and replace /'s with _'s. If set, this enables the standard SavedModel
-// classification, regression, and prediction signatures to be used as
-// uppercase INPUTS and OUTPUTS tags for streams.
+// If options.convert_signature_to_tags() is set, will convert letters to
+// uppercase and replace /'s and -'s with _'s. This enables the standard
+// SavedModel classification, regression, and prediction signatures to be used
+// as uppercase INPUTS and OUTPUTS tags for streams and supports other common
+// patterns.
 const std::string MaybeConvertSignatureToTag(
     const std::string& name,
     const TensorFlowSessionFromSavedModelCalculatorOptions& options) {
@@ -63,7 +64,8 @@ const std::string MaybeConvertSignatureToTag(
     output.resize(name.length());
     std::transform(name.begin(), name.end(), output.begin(),
                    [](unsigned char c) { return std::toupper(c); });
-    output = absl::Substitute(output, "/", "_");
+    output = absl::StrReplaceAll(output, {{"/", "_"}});
+    output = absl::StrReplaceAll(output, {{"-", "_"}});
     return output;
   } else {
     return name;

mediapipe/calculators/tensorflow/tensorflow_session_from_saved_model_calculator.proto

@@ -32,8 +32,8 @@ message TensorFlowSessionFromSavedModelCalculatorOptions {
   // The name of the generic signature to load into the mapping from tags to
   // tensor names.
   optional string signature_name = 2 [default = "serving_default"];
-  // Whether to convert the signature keys to uppercase and switch /'s to
-  // _'s, which enables standard signatures to be used as Tags.
+  // Whether to convert the signature keys to uppercase as well as switch /'s
+  // and -'s to _'s, which enables common signatures to be used as Tags.
   optional bool convert_signature_to_tags = 3 [default = true];
   // If true, saved_model_path can have multiple exported models in
   // subdirectories saved_model_path/%08d and the alphabetically last (i.e.,

mediapipe/calculators/tensorflow/tensorflow_session_from_saved_model_calculator_test.cc

@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "absl/strings/substitute.h"
+#include "absl/strings/str_replace.h"
 #include "mediapipe/calculators/tensorflow/tensorflow_session.h"
 #include "mediapipe/calculators/tensorflow/tensorflow_session_from_saved_model_calculator.pb.h"
 #include "mediapipe/framework/calculator.pb.h"

mediapipe/calculators/tensorflow/tensorflow_session_from_saved_model_generator.cc

@@ -17,7 +17,7 @@
 #if !defined(__ANDROID__)
 #include "mediapipe/framework/port/file_helpers.h"
 #endif
-#include "absl/strings/substitute.h"
+#include "absl/strings/str_replace.h"
 #include "mediapipe/calculators/tensorflow/tensorflow_session.h"
 #include "mediapipe/calculators/tensorflow/tensorflow_session_from_saved_model_generator.pb.h"
 #include "mediapipe/framework/deps/file_path.h"
@@ -53,10 +53,11 @@ static constexpr char kStringSavedModelPath[] = "STRING_SAVED_MODEL_PATH";
 #endif
 }
 
-// If options.convert_signature_to_tags() will convert letters to uppercase
-// and replace /'s with _'s. If set, this enables the standard SavedModel
-// classification, regression, and prediction signatures to be used as
-// uppercase INPUTS and OUTPUTS tags for streams.
+// If options.convert_signature_to_tags() is set, will convert letters to
+// uppercase and replace /'s and -'s with _'s. This enables the standard
+// SavedModel classification, regression, and prediction signatures to be used
+// as uppercase INPUTS and OUTPUTS tags for streams and supports other common
+// patterns.
 const std::string MaybeConvertSignatureToTag(
     const std::string& name,
     const TensorFlowSessionFromSavedModelGeneratorOptions& options) {
@@ -65,7 +66,8 @@ const std::string MaybeConvertSignatureToTag(
     output.resize(name.length());
     std::transform(name.begin(), name.end(), output.begin(),
                    [](unsigned char c) { return std::toupper(c); });
-    output = absl::Substitute(output, "/", "_");
+    output = absl::StrReplaceAll(output, {{"/", "_"}});
+    output = absl::StrReplaceAll(output, {{"-", "_"}});
     return output;
   } else {
     return name;

mediapipe/calculators/tensorflow/tensorflow_session_from_saved_model_generator.proto

@@ -32,8 +32,8 @@ message TensorFlowSessionFromSavedModelGeneratorOptions {
   // The name of the generic signature to load into the mapping from tags to
   // tensor names.
   optional string signature_name = 2 [default = "serving_default"];
-  // Whether to convert the signature keys to uppercase and switch /'s to
-  // _'s, which enables standard signatures to be used as Tags.
+  // Whether to convert the signature keys to uppercase as well as switch /'s
+  // and -'s to _'s, which enables common signatures to be used as Tags.
   optional bool convert_signature_to_tags = 3 [default = true];
   // If true, saved_model_path can have multiple exported models in
   // subdirectories saved_model_path/%08d and the alphabetically last (i.e.,

mediapipe/calculators/tensorflow/tensorflow_session_from_saved_model_generator_test.cc

@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "absl/strings/substitute.h"
+#include "absl/strings/str_replace.h"
 #include "mediapipe/calculators/tensorflow/tensorflow_session.h"
 #include "mediapipe/calculators/tensorflow/tensorflow_session_from_saved_model_generator.pb.h"
 #include "mediapipe/framework/calculator_framework.h"

mediapipe/calculators/tflite/BUILD

@@ -222,9 +222,11 @@ cc_library(
     deps = [
         ":util",
         ":tflite_inference_calculator_cc_proto",
+        "@com_google_absl//absl/memory",
         "//mediapipe/framework:calculator_framework",
         "//mediapipe/util:resource_util",
         "@org_tensorflow//tensorflow/lite:framework",
+        "@org_tensorflow//tensorflow/lite/delegates/xnnpack:xnnpack_delegate",
         "@org_tensorflow//tensorflow/lite/kernels:builtin_ops",
         "//mediapipe/framework/stream_handler:fixed_size_input_stream_handler",
         "//mediapipe/framework/port:ret_check",
@@ -254,6 +256,10 @@ cc_library(
         "//mediapipe:android": [
             "@org_tensorflow//tensorflow/lite/delegates/nnapi:nnapi_delegate",
         ],
+    }) + select({
+        "//conditions:default": [
+            "//mediapipe/util:cpu_util",
+        ],
     }),
     alwayslink = 1,
 )
@@ -308,6 +314,20 @@ cc_library(
     alwayslink = 1,
 )
 
+cc_library(
+    name = "tflite_model_calculator",
+    srcs = ["tflite_model_calculator.cc"],
+    visibility = ["//visibility:public"],
+    deps = [
+        ":util",
+        "//mediapipe/framework:calculator_framework",
+        "//mediapipe/framework:packet",
+        "//mediapipe/framework/port:ret_check",
+        "@org_tensorflow//tensorflow/lite:framework",
+    ],
+    alwayslink = 1,
+)
+
 cc_library(
     name = "tflite_tensors_to_segmentation_calculator",
     srcs = ["tflite_tensors_to_segmentation_calculator.cc"],
@@ -431,7 +451,7 @@ cc_library(
         "//mediapipe:android": [
             "//mediapipe/util/android/file/base",
         ],
-        "//mediapipe:apple": [
+        "//mediapipe:ios": [
             "//mediapipe/util/android/file/base",
         ],
         "//mediapipe:macos": [
@@ -478,6 +498,9 @@ cc_test(
     deps = [
         ":tflite_inference_calculator",
         ":tflite_inference_calculator_cc_proto",
+        ":tflite_model_calculator",
+        "//mediapipe/calculators/core:constant_side_packet_calculator",
+        "//mediapipe/calculators/util:local_file_contents_calculator",
         "//mediapipe/framework:calculator_framework",
         "//mediapipe/framework:calculator_runner",
         "//mediapipe/framework/deps:file_path",
@@ -485,6 +508,9 @@ cc_test(
         "//mediapipe/framework/port:integral_types",
         "//mediapipe/framework/port:parse_text_proto",
         "//mediapipe/framework/tool:validate_type",
+        "@com_google_absl//absl/memory",
+        "@com_google_absl//absl/strings",
+        "@com_google_absl//absl/types:optional",
         "@org_tensorflow//tensorflow/lite:framework",
         "@org_tensorflow//tensorflow/lite/kernels:builtin_ops",
     ],
@@ -510,3 +536,19 @@ cc_test(
         "@org_tensorflow//tensorflow/lite/kernels:builtin_ops",
     ],
 )
+
+cc_test(
+    name = "tflite_model_calculator_test",
+    srcs = ["tflite_model_calculator_test.cc"],
+    data = ["testdata/add.bin"],
+    deps = [
+        ":tflite_model_calculator",
+        "//mediapipe/calculators/core:constant_side_packet_calculator",
+        "//mediapipe/calculators/util:local_file_contents_calculator",
+        "//mediapipe/framework:calculator_framework",
+        "//mediapipe/framework:calculator_runner",
+        "//mediapipe/framework/port:gtest_main",
+        "//mediapipe/framework/port:parse_text_proto",
+        "@org_tensorflow//tensorflow/lite:framework",
+    ],
+)

mediapipe/calculators/tflite/tflite_inference_calculator.cc

@@ -17,10 +17,16 @@
 #include <string>
 #include <vector>
 
+#include "absl/memory/memory.h"
 #include "mediapipe/calculators/tflite/tflite_inference_calculator.pb.h"
 #include "mediapipe/calculators/tflite/util.h"
 #include "mediapipe/framework/calculator_framework.h"
 #include "mediapipe/framework/port/ret_check.h"
+
+#if !defined(__EMSCRIPTEN__)
+#include "mediapipe/util/cpu_util.h"
+#endif  // !__EMSCRIPTEN__
+
 #include "mediapipe/util/resource_util.h"
 #include "tensorflow/lite/error_reporter.h"
 #include "tensorflow/lite/interpreter.h"
@@ -50,7 +56,7 @@
 #include "tensorflow/lite/delegates/gpu/metal_delegate.h"
 #include "tensorflow/lite/delegates/gpu/metal_delegate_internal.h"
 #endif  // iOS
-
+#include "tensorflow/lite/delegates/xnnpack/xnnpack_delegate.h"
 #if defined(MEDIAPIPE_ANDROID)
 #include "tensorflow/lite/delegates/nnapi/nnapi_delegate.h"
 #endif  // ANDROID
@@ -113,6 +119,23 @@ struct GPUData {
 };
 #endif
 
+// Returns number of threads to configure XNNPACK delegate with.
+// (Equal to user provided value if specified.  Otherwise, it returns number of
+// high cores (hard-coded to 1 for __EMSCRIPTEN__))
+int GetXnnpackNumThreads(
+    const mediapipe::TfLiteInferenceCalculatorOptions& opts) {
+  static constexpr int kDefaultNumThreads = -1;
+  if (opts.has_delegate() && opts.delegate().has_xnnpack() &&
+      opts.delegate().xnnpack().num_threads() != kDefaultNumThreads) {
+    return opts.delegate().xnnpack().num_threads();
+  }
+#if !defined(__EMSCRIPTEN__)
+  return InferHigherCoreIds().size();
+#else
+  return 1;
+#endif  // !__EMSCRIPTEN__
+}
+
 // Calculator Header Section
 
 // Runs inference on the provided input TFLite tensors and TFLite model.
@@ -139,6 +162,9 @@ struct GPUData {
 // Input side packet:
 //  CUSTOM_OP_RESOLVER (optional) - Use a custom op resolver,
 //                                  instead of the builtin one.
+//  MODEL (optional) - Use to specify TfLite model
+//                     (std::unique_ptr<tflite::FlatBufferModel,
+//                       std::function<void(tflite::FlatBufferModel*)>>)
 //
 // Example use:
 // node {
@@ -148,7 +174,21 @@ struct GPUData {
 //   options: {
 //     [mediapipe.TfLiteInferenceCalculatorOptions.ext] {
 //       model_path: "modelname.tflite"
-//       use_gpu: true
+//       delegate { gpu {} }
+//     }
+//   }
+// }
+//
+// or
+//
+// node {
+//   calculator: "TfLiteInferenceCalculator"
+//   input_stream: "TENSORS:tensor_image"
+//   input_side_packet: "MODEL:model"
+//   output_stream: "TENSORS:tensors"
+//   options: {
+//     [mediapipe.TfLiteInferenceCalculatorOptions.ext] {
+//       delegate { gpu {} }
 //     }
 //   }
 // }
@@ -163,6 +203,12 @@ struct GPUData {
 //
 class TfLiteInferenceCalculator : public CalculatorBase {
  public:
+  using TfLiteDelegatePtr =
+      std::unique_ptr<TfLiteDelegate, std::function<void(TfLiteDelegate*)>>;
+  using TfLiteModelPtr =
+      std::unique_ptr<tflite::FlatBufferModel,
+                      std::function<void(tflite::FlatBufferModel*)>>;
+
   static ::mediapipe::Status GetContract(CalculatorContract* cc);
 
   ::mediapipe::Status Open(CalculatorContext* cc) override;
@@ -170,13 +216,13 @@ class TfLiteInferenceCalculator : public CalculatorBase {
   ::mediapipe::Status Close(CalculatorContext* cc) override;
 
  private:
-  ::mediapipe::Status LoadOptions(CalculatorContext* cc);
   ::mediapipe::Status LoadModel(CalculatorContext* cc);
+  ::mediapipe::StatusOr<Packet> GetModelAsPacket(const CalculatorContext& cc);
   ::mediapipe::Status LoadDelegate(CalculatorContext* cc);
 
+  Packet model_packet_;
   std::unique_ptr<tflite::Interpreter> interpreter_;
-  std::unique_ptr<tflite::FlatBufferModel> model_;
-  TfLiteDelegate* delegate_ = nullptr;
+  TfLiteDelegatePtr delegate_;
 
 #if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
   mediapipe::GlCalculatorHelper gpu_helper_;
@@ -195,7 +241,6 @@ class TfLiteInferenceCalculator : public CalculatorBase {
       edgetpu::EdgeTpuManager::GetSingleton()->OpenDevice();
 #endif
 
-  std::string model_path_ = "";
   bool gpu_inference_ = false;
   bool gpu_input_ = false;
   bool gpu_output_ = false;
@@ -212,12 +257,22 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
   RET_CHECK(cc->Outputs().HasTag("TENSORS") ^
             cc->Outputs().HasTag("TENSORS_GPU"));
 
-  bool use_gpu = false;
+  const auto& options =
+      cc->Options<::mediapipe::TfLiteInferenceCalculatorOptions>();
+  RET_CHECK(!options.model_path().empty() ^
+            cc->InputSidePackets().HasTag("MODEL"))
+      << "Either model as side packet or model path in options is required.";
+
+  bool use_gpu =
+      options.has_delegate() ? options.delegate().has_gpu() : options.use_gpu();
 
   if (cc->Inputs().HasTag("TENSORS"))
     cc->Inputs().Tag("TENSORS").Set<std::vector<TfLiteTensor>>();
 #if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__)
   if (cc->Inputs().HasTag("TENSORS_GPU")) {
+    RET_CHECK(!options.has_delegate() || options.delegate().has_gpu())
+        << "GPU input is compatible with GPU delegate only.";
+
     cc->Inputs().Tag("TENSORS_GPU").Set<std::vector<GpuTensor>>();
     use_gpu |= true;
   }
@@ -227,6 +282,9 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
     cc->Outputs().Tag("TENSORS").Set<std::vector<TfLiteTensor>>();
 #if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__)
   if (cc->Outputs().HasTag("TENSORS_GPU")) {
+    RET_CHECK(!options.has_delegate() || options.delegate().has_gpu())
+        << "GPU output is compatible with GPU delegate only.";
+
     cc->Outputs().Tag("TENSORS_GPU").Set<std::vector<GpuTensor>>();
     use_gpu |= true;
   }
@@ -237,10 +295,9 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
         .Tag("CUSTOM_OP_RESOLVER")
         .Set<tflite::ops::builtin::BuiltinOpResolver>();
   }
-
-  const auto& options =
-      cc->Options<::mediapipe::TfLiteInferenceCalculatorOptions>();
-  use_gpu |= options.use_gpu();
+  if (cc->InputSidePackets().HasTag("MODEL")) {
+    cc->InputSidePackets().Tag("MODEL").Set<TfLiteModelPtr>();
+  }
 
   if (use_gpu) {
 #if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
@@ -259,7 +316,9 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
 ::mediapipe::Status TfLiteInferenceCalculator::Open(CalculatorContext* cc) {
   cc->SetOffset(TimestampDiff(0));
 
-  MP_RETURN_IF_ERROR(LoadOptions(cc));
+  const auto& options =
+      cc->Options<::mediapipe::TfLiteInferenceCalculatorOptions>();
+  gpu_inference_ = options.use_gpu();
 
   if (cc->Inputs().HasTag("TENSORS_GPU")) {
 #if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__)
@@ -454,7 +513,7 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
     if (gpu_inference_) {
 #if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
       MP_RETURN_IF_ERROR(gpu_helper_.RunInGlContext([this]() -> Status {
-        TfLiteGpuDelegateDelete(delegate_);
+        delegate_ = nullptr;
         for (int i = 0; i < gpu_data_in_.size(); ++i) {
           gpu_data_in_[i].reset();
         }
@@ -464,7 +523,7 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
         return ::mediapipe::OkStatus();
       }));
 #elif defined(MEDIAPIPE_IOS)
-      TFLGpuDelegateDelete(delegate_);
+      delegate_ = nullptr;
       for (int i = 0; i < gpu_data_in_.size(); ++i) {
         gpu_data_in_[i].reset();
       }
@@ -472,8 +531,9 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
         gpu_data_out_[i].reset();
       }
 #endif
+    } else {
+      delegate_ = nullptr;
     }
-    delegate_ = nullptr;
   }
 #if defined(MEDIAPIPE_EDGE_TPU)
   edgetpu_context_.reset();
@@ -483,33 +543,10 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
 
 // Calculator Auxiliary Section
 
-::mediapipe::Status TfLiteInferenceCalculator::LoadOptions(
-    CalculatorContext* cc) {
-  // Get calculator options specified in the graph.
-  const auto& options =
-      cc->Options<::mediapipe::TfLiteInferenceCalculatorOptions>();
-
-  // Get model name.
-  if (!options.model_path().empty()) {
-    std::string model_path = options.model_path();
-
-    ASSIGN_OR_RETURN(model_path_, mediapipe::PathToResourceAsFile(model_path));
-  } else {
-    LOG(ERROR) << "Must specify path to TFLite model.";
-    return ::mediapipe::Status(::mediapipe::StatusCode::kNotFound,
-                               "Must specify path to TFLite model.");
-  }
-
-  // Get execution modes.
-  gpu_inference_ = options.use_gpu();
-
-  return ::mediapipe::OkStatus();
-}
-
 ::mediapipe::Status TfLiteInferenceCalculator::LoadModel(
     CalculatorContext* cc) {
-  model_ = tflite::FlatBufferModel::BuildFromFile(model_path_.c_str());
-  RET_CHECK(model_);
+  ASSIGN_OR_RETURN(model_packet_, GetModelAsPacket(*cc));
+  const auto& model = *model_packet_.Get<TfLiteModelPtr>();
 
   tflite::ops::builtin::BuiltinOpResolver op_resolver;
   if (cc->InputSidePackets().HasTag("CUSTOM_OP_RESOLVER")) {
@@ -519,15 +556,19 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
   }
 #if defined(MEDIAPIPE_EDGE_TPU)
   interpreter_ =
-      BuildEdgeTpuInterpreter(*model_, &op_resolver, edgetpu_context_.get());
+      BuildEdgeTpuInterpreter(model, &op_resolver, edgetpu_context_.get());
 #else
-  tflite::InterpreterBuilder(*model_, op_resolver)(&interpreter_);
+  tflite::InterpreterBuilder(model, op_resolver)(&interpreter_);
 #endif  // MEDIAPIPE_EDGE_TPU
 
   RET_CHECK(interpreter_);
 
-#if defined(__EMSCRIPTEN__)
+#if defined(__EMSCRIPTEN__) || defined(MEDIAPIPE_EDGE_TPU)
   interpreter_->SetNumThreads(1);
+#else
+  interpreter_->SetNumThreads(
+      cc->Options<mediapipe::TfLiteInferenceCalculatorOptions>()
+          .cpu_num_thread());
 #endif  // __EMSCRIPTEN__
 
   if (gpu_output_) {
@@ -543,22 +584,77 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
   return ::mediapipe::OkStatus();
 }
 
+::mediapipe::StatusOr<Packet> TfLiteInferenceCalculator::GetModelAsPacket(
+    const CalculatorContext& cc) {
+  const auto& options =
+      cc.Options<mediapipe::TfLiteInferenceCalculatorOptions>();
+  if (!options.model_path().empty()) {
+    std::string model_path = options.model_path();
+
+    ASSIGN_OR_RETURN(model_path, mediapipe::PathToResourceAsFile(model_path));
+
+    auto model = tflite::FlatBufferModel::BuildFromFile(model_path.c_str());
+    RET_CHECK(model) << "Failed to load model from path.";
+    return MakePacket<TfLiteModelPtr>(TfLiteModelPtr(
+        model.release(), [](tflite::FlatBufferModel* model) { delete model; }));
+  }
+  if (cc.InputSidePackets().HasTag("MODEL")) {
+    return cc.InputSidePackets().Tag("MODEL");
+  }
+  return ::mediapipe::Status(
+      ::mediapipe::StatusCode::kNotFound,
+      "Must specify TFLite model as path or loaded model.");
+}
+
 ::mediapipe::Status TfLiteInferenceCalculator::LoadDelegate(
     CalculatorContext* cc) {
-#if defined(MEDIAPIPE_ANDROID)
+  const auto& calculator_opts =
+      cc->Options<mediapipe::TfLiteInferenceCalculatorOptions>();
+  if (calculator_opts.has_delegate() &&
+      calculator_opts.delegate().has_tflite()) {
+    // Default tflite inference requeqsted - no need to modify graph.
+    return ::mediapipe::OkStatus();
+  }
+
   if (!gpu_inference_) {
-    if (cc->Options<mediapipe::TfLiteInferenceCalculatorOptions>()
-            .use_nnapi()) {
+#if defined(MEDIAPIPE_ANDROID)
+    const bool nnapi_requested = calculator_opts.has_delegate()
+                                     ? calculator_opts.delegate().has_nnapi()
+                                     : calculator_opts.use_nnapi();
+    if (nnapi_requested) {
       // Attempt to use NNAPI.
       // If not supported, the default CPU delegate will be created and used.
       interpreter_->SetAllowFp16PrecisionForFp32(1);
-      delegate_ = tflite::NnApiDelegate();
-      RET_CHECK_EQ(interpreter_->ModifyGraphWithDelegate(delegate_), kTfLiteOk);
+      delegate_ =
+          TfLiteDelegatePtr(tflite::NnApiDelegate(), [](TfLiteDelegate*) {
+            // No need to free according to tflite::NnApiDelegate()
+            // documentation.
+          });
+      RET_CHECK_EQ(interpreter_->ModifyGraphWithDelegate(delegate_.get()),
+                   kTfLiteOk);
+      return ::mediapipe::OkStatus();
     }
+#endif  // MEDIAPIPE_ANDROID
+
+#if defined(__EMSCRIPTEN__)
+    const bool xnnpack_requested = true;
+#else
+    const bool xnnpack_requested = calculator_opts.has_delegate() &&
+                                   calculator_opts.delegate().has_xnnpack();
+#endif  // __EMSCRIPTEN__
+
+    if (xnnpack_requested) {
+      TfLiteXNNPackDelegateOptions xnnpack_opts{};
+      xnnpack_opts.num_threads = GetXnnpackNumThreads(calculator_opts);
+      delegate_ = TfLiteDelegatePtr(TfLiteXNNPackDelegateCreate(&xnnpack_opts),
+                                    &TfLiteXNNPackDelegateDelete);
+      RET_CHECK_EQ(interpreter_->ModifyGraphWithDelegate(delegate_.get()),
+                   kTfLiteOk);
+    }
+
     // Return, no need for GPU delegate below.
     return ::mediapipe::OkStatus();
   }
-#endif  // ANDROID
 
 #if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
   // Configure and create the delegate.
@@ -568,14 +664,16 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
       TFLITE_GL_OBJECT_TYPE_FASTEST;
   options.compile_options.dynamic_batch_enabled = 0;
   options.compile_options.inline_parameters = 1;
-  if (!delegate_) delegate_ = TfLiteGpuDelegateCreate(&options);
+  if (!delegate_)
+    delegate_ = TfLiteDelegatePtr(TfLiteGpuDelegateCreate(&options),
+                                  &TfLiteGpuDelegateDelete);
 
   if (gpu_input_) {
     // Get input image sizes.
     const auto& input_indices = interpreter_->inputs();
     gpu_data_in_.resize(input_indices.size());
     for (int i = 0; i < input_indices.size(); ++i) {
-      const TfLiteTensor* tensor = interpreter_->tensor(input_indices[0]);
+      const TfLiteTensor* tensor = interpreter_->tensor(input_indices[i]);
       gpu_data_in_[i] = absl::make_unique<GPUData>();
       gpu_data_in_[i]->elements = 1;
       for (int d = 0; d < tensor->dims->size; ++d) {
@@ -586,7 +684,7 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
           ::tflite::gpu::gl::CreateReadWriteShaderStorageBuffer<float>(
               gpu_data_in_[i]->elements, &gpu_data_in_[i]->buffer));
       RET_CHECK_EQ(TfLiteGpuDelegateBindBufferToTensor(
-                       delegate_, gpu_data_in_[i]->buffer.id(),
+                       delegate_.get(), gpu_data_in_[i]->buffer.id(),
                        interpreter_->inputs()[i]),
                    kTfLiteOk);
     }
@@ -609,15 +707,16 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
     for (int i = 0; i < gpu_data_out_.size(); ++i) {
       RET_CHECK_CALL(CreateReadWriteShaderStorageBuffer<float>(
           gpu_data_out_[i]->elements, &gpu_data_out_[i]->buffer));
-      RET_CHECK_EQ(
-          TfLiteGpuDelegateBindBufferToTensor(
-              delegate_, gpu_data_out_[i]->buffer.id(), output_indices[i]),
-          kTfLiteOk);
+      RET_CHECK_EQ(TfLiteGpuDelegateBindBufferToTensor(
+                       delegate_.get(), gpu_data_out_[i]->buffer.id(),
+                       output_indices[i]),
+                   kTfLiteOk);
     }
   }
 
   // Must call this last.
-  RET_CHECK_EQ(interpreter_->ModifyGraphWithDelegate(delegate_), kTfLiteOk);
+  RET_CHECK_EQ(interpreter_->ModifyGraphWithDelegate(delegate_.get()),
+               kTfLiteOk);
 #endif  // OpenGL
 
 #if defined(MEDIAPIPE_IOS)
@@ -626,7 +725,9 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
   TFLGpuDelegateOptions options;
   options.allow_precision_loss = true;
   options.wait_type = TFLGpuDelegateWaitType::TFLGpuDelegateWaitTypePassive;
-  if (!delegate_) delegate_ = TFLGpuDelegateCreate(&options);
+  if (!delegate_)
+    delegate_ = TfLiteDelegatePtr(TFLGpuDelegateCreate(&options),
+                                  &TFLGpuDelegateDelete);
   id<MTLDevice> device = gpu_helper_.mtlDevice;
 
   if (gpu_input_) {
@@ -678,10 +779,12 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
       gpu_data_in_[i]->buffer =
           [device newBufferWithLength:gpu_data_in_[i]->elements * kHalfSize
                               options:MTLResourceStorageModeShared];
-      RET_CHECK_EQ(interpreter_->ModifyGraphWithDelegate(delegate_), kTfLiteOk);
-      RET_CHECK_EQ(TFLGpuDelegateBindMetalBufferToTensor(
-                       delegate_, input_indices[i], gpu_data_in_[i]->buffer),
-                   true);
+      RET_CHECK_EQ(interpreter_->ModifyGraphWithDelegate(delegate_.get()),
+                   kTfLiteOk);
+      RET_CHECK_EQ(
+          TFLGpuDelegateBindMetalBufferToTensor(
+              delegate_.get(), input_indices[i], gpu_data_in_[i]->buffer),
+          true);
     }
   }
   if (gpu_output_) {
@@ -725,9 +828,10 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
       gpu_data_out_[i]->buffer =
           [device newBufferWithLength:gpu_data_out_[i]->elements * kHalfSize
                               options:MTLResourceStorageModeShared];
-      RET_CHECK_EQ(TFLGpuDelegateBindMetalBufferToTensor(
-                       delegate_, output_indices[i], gpu_data_out_[i]->buffer),
-                   true);
+      RET_CHECK_EQ(
+          TFLGpuDelegateBindMetalBufferToTensor(
+              delegate_.get(), output_indices[i], gpu_data_out_[i]->buffer),
+          true);
     }
 
     // Create converter for GPU output.

mediapipe/calculators/tflite/tflite_inference_calculator.proto

@@ -27,7 +27,7 @@ import "mediapipe/framework/calculator.proto";
 //   options {
 //     [mediapipe.TfLiteInferenceCalculatorOptions.ext] {
 //       model_path: "model.tflite"
-//       use_gpu: true
+//       delegate { gpu {} }
 //     }
 //   }
 // }
@@ -37,6 +37,28 @@ message TfLiteInferenceCalculatorOptions {
     optional TfLiteInferenceCalculatorOptions ext = 233867213;
   }
 
+  message Delegate {
+    // Default inference provided by tflite.
+    message TfLite {}
+    // Delegate to run GPU inference depending on the device.
+    // (Can use OpenGl, OpenCl, Metal depending on the device.)
+    message Gpu {}
+    // Android only.
+    message Nnapi {}
+    message Xnnpack {
+      // Number of threads for XNNPACK delegate. (By default, calculator tries
+      // to choose optimal number of threads depending on the device.)
+      optional int32 num_threads = 1 [default = -1];
+    }
+
+    oneof delegate {
+      TfLite tflite = 1;
+      Gpu gpu = 2;
+      Nnapi nnapi = 3;
+      Xnnpack xnnpack = 4;
+    }
+  }
+
   // Path to the TF Lite model (ex: /path/to/modelname.tflite).
   // On mobile, this is generally just modelname.tflite.
   optional string model_path = 1;
@@ -44,10 +66,22 @@ message TfLiteInferenceCalculatorOptions {
   // Whether the TF Lite GPU or CPU backend should be used. Effective only when
   // input tensors are on CPU. For input tensors on GPU, GPU backend is always
   // used.
-  optional bool use_gpu = 2 [default = false];
+  // DEPRECATED: configure "delegate" instead.
+  optional bool use_gpu = 2 [deprecated = true, default = false];
 
   // Android only. When true, an NNAPI delegate will be used for inference.
   // If NNAPI is not available, then the default CPU delegate will be used
   // automatically.
-  optional bool use_nnapi = 3 [default = false];
+  // DEPRECATED: configure "delegate" instead.
+  optional bool use_nnapi = 3 [deprecated = true, default = false];
+
+  // The number of threads available to the interpreter. Effective only when
+  // input tensors are on CPU and 'use_gpu' is false.
+  optional int32 cpu_num_thread = 4 [default = -1];
+
+  // TfLite delegate to run inference.
+  // NOTE: calculator is free to choose delegate if not specified explicitly.
+  // NOTE: use_gpu/use_nnapi are ignored if specified. (Delegate takes
+  // precedence over use_* deprecated options.)
+  optional Delegate delegate = 5;
 }

mediapipe/calculators/tflite/tflite_inference_calculator_test.cc

@@ -16,6 +16,8 @@
 #include <string>
 #include <vector>
 
+#include "absl/strings/str_replace.h"
+#include "absl/strings/string_view.h"
 #include "mediapipe/calculators/tflite/tflite_inference_calculator.pb.h"
 #include "mediapipe/framework/calculator_framework.h"
 #include "mediapipe/framework/calculator_runner.h"
@@ -39,13 +41,7 @@ namespace mediapipe {
 
 using ::tflite::Interpreter;
 
-class TfLiteInferenceCalculatorTest : public ::testing::Test {
- protected:
-  std::unique_ptr<CalculatorRunner> runner_ = nullptr;
-};
-
-// Tests a simple add model that adds an input tensor to itself.
-TEST_F(TfLiteInferenceCalculatorTest, SmokeTest) {
+void DoSmokeTest(const std::string& graph_proto) {
   const int width = 8;
   const int height = 8;
   const int channels = 3;
@@ -75,21 +71,7 @@ TEST_F(TfLiteInferenceCalculatorTest, SmokeTest) {
 
   // Prepare single calculator graph to and wait for packets.
   CalculatorGraphConfig graph_config =
-      ::mediapipe::ParseTextProtoOrDie<CalculatorGraphConfig>(
-          R"(
-            input_stream: "tensor_in"
-            node {
-              calculator: "TfLiteInferenceCalculator"
-              input_stream: "TENSORS:tensor_in"
-              output_stream: "TENSORS:tensor_out"
-              options {
-                [mediapipe.TfLiteInferenceCalculatorOptions.ext] {
-                  use_gpu: false
-                  model_path: "mediapipe/calculators/tflite/testdata/add.bin"
-                }
-              }
-            }
-          )");
+      ParseTextProtoOrDie<CalculatorGraphConfig>(graph_proto);
   std::vector<Packet> output_packets;
   tool::AddVectorSink("tensor_out", &graph_config, &output_packets);
   CalculatorGraph graph(graph_config);
@@ -120,4 +102,72 @@ TEST_F(TfLiteInferenceCalculatorTest, SmokeTest) {
   MP_ASSERT_OK(graph.WaitUntilDone());
 }
 
+// Tests a simple add model that adds an input tensor to itself.
+TEST(TfLiteInferenceCalculatorTest, SmokeTest) {
+  std::string graph_proto = R"(
+    input_stream: "tensor_in"
+    node {
+      calculator: "TfLiteInferenceCalculator"
+      input_stream: "TENSORS:tensor_in"
+      output_stream: "TENSORS:tensor_out"
+      options {
+        [mediapipe.TfLiteInferenceCalculatorOptions.ext] {
+          model_path: "mediapipe/calculators/tflite/testdata/add.bin"
+          $delegate
+        }
+      }
+    }
+  )";
+  DoSmokeTest(
+      /*graph_proto=*/absl::StrReplaceAll(graph_proto, {{"$delegate", ""}}));
+  DoSmokeTest(/*graph_proto=*/absl::StrReplaceAll(
+      graph_proto, {{"$delegate", "delegate { tflite {} }"}}));
+  DoSmokeTest(/*graph_proto=*/absl::StrReplaceAll(
+      graph_proto, {{"$delegate", "delegate { xnnpack {} }"}}));
+  DoSmokeTest(/*graph_proto=*/absl::StrReplaceAll(
+      graph_proto,
+      {{"$delegate", "delegate { xnnpack { num_threads: 10 } }"}}));
+}
+
+TEST(TfLiteInferenceCalculatorTest, SmokeTest_ModelAsInputSidePacket) {
+  std::string graph_proto = R"(
+    input_stream: "tensor_in"
+
+    node {
+      calculator: "ConstantSidePacketCalculator"
+      output_side_packet: "PACKET:model_path"
+      options: {
+        [mediapipe.ConstantSidePacketCalculatorOptions.ext]: {
+          packet { string_value: "mediapipe/calculators/tflite/testdata/add.bin" }
+        }
+      }
+    }
+
+    node {
+      calculator: "LocalFileContentsCalculator"
+      input_side_packet: "FILE_PATH:model_path"
+      output_side_packet: "CONTENTS:model_blob"
+    }
+
+    node {
+      calculator: "TfLiteModelCalculator"
+      input_side_packet: "MODEL_BLOB:model_blob"
+      output_side_packet: "MODEL:model"
+    }
+
+    node {
+      calculator: "TfLiteInferenceCalculator"
+      input_stream: "TENSORS:tensor_in"
+      output_stream: "TENSORS:tensor_out"
+      input_side_packet: "MODEL:model"
+      options {
+        [mediapipe.TfLiteInferenceCalculatorOptions.ext] {
+          use_gpu: false
+        }
+      }
+    }
+  )";
+  DoSmokeTest(graph_proto);
+}
+
 }  // namespace mediapipe

mediapipe/calculators/tflite/tflite_model_calculator.cc

@@ -0,0 +1,86 @@
+// Copyright 2020 The MediaPipe Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <functional>
+#include <memory>
+#include <string>
+
+#include "mediapipe/framework/calculator_framework.h"
+#include "mediapipe/framework/packet.h"
+#include "mediapipe/framework/port/ret_check.h"
+#include "tensorflow/lite/model.h"
+
+namespace mediapipe {
+
+// Loads TfLite model from model blob specified as input side packet and outputs
+// corresponding side packet.
+//
+// Input side packets:
+//   MODEL_BLOB - TfLite model blob/file-contents (std::string). You can read
+//                model blob from file (using whatever APIs you have) and pass
+//                it to the graph as input side packet or you can use some of
+//                calculators like LocalFileContentsCalculator to get model
+//                blob and use it as input here.
+//
+// Output side packets:
+//   MODEL - TfLite model. (std::unique_ptr<tflite::FlatBufferModel,
+//           std::function<void(tflite::FlatBufferModel*)>>)
+//
+// Example use:
+//
+// node {
+//   calculator: "TfLiteModelCalculator"
+//   input_side_packet: "MODEL_BLOB:model_blob"
+//   output_side_packet: "MODEL:model"
+// }
+//
+class TfLiteModelCalculator : public CalculatorBase {
+ public:
+  using TfLiteModelPtr =
+      std::unique_ptr<tflite::FlatBufferModel,
+                      std::function<void(tflite::FlatBufferModel*)>>;
+
+  static ::mediapipe::Status GetContract(CalculatorContract* cc) {
+    cc->InputSidePackets().Tag("MODEL_BLOB").Set<std::string>();
+    cc->OutputSidePackets().Tag("MODEL").Set<TfLiteModelPtr>();
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Open(CalculatorContext* cc) override {
+    const Packet& model_packet = cc->InputSidePackets().Tag("MODEL_BLOB");
+    const std::string& model_blob = model_packet.Get<std::string>();
+    std::unique_ptr<tflite::FlatBufferModel> model =
+        tflite::FlatBufferModel::BuildFromBuffer(model_blob.data(),
+                                                 model_blob.size());
+    RET_CHECK(model) << "Failed to load TfLite model from blob.";
+
+    cc->OutputSidePackets().Tag("MODEL").Set(
+        MakePacket<TfLiteModelPtr>(TfLiteModelPtr(
+            model.release(), [model_packet](tflite::FlatBufferModel* model) {
+              // Keeping model_packet in order to keep underlying model blob
+              // which can be released only after TfLite model is not needed
+              // anymore (deleted).
+              delete model;
+            })));
+
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Process(CalculatorContext* cc) override {
+    return ::mediapipe::OkStatus();
+  }
+};
+REGISTER_CALCULATOR(TfLiteModelCalculator);
+
+}  // namespace mediapipe

mediapipe/calculators/tflite/tflite_model_calculator_test.cc

@@ -0,0 +1,88 @@
+// Copyright 2020 The MediaPipe Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <memory>
+
+#include "mediapipe/framework/calculator_framework.h"
+#include "mediapipe/framework/calculator_runner.h"
+#include "mediapipe/framework/port/gmock.h"
+#include "mediapipe/framework/port/gtest.h"
+#include "mediapipe/framework/port/parse_text_proto.h"
+#include "mediapipe/framework/port/status_matchers.h"  // NOLINT
+#include "tensorflow/lite/model.h"
+
+namespace mediapipe {
+
+TEST(TfLiteModelCalculatorTest, SmokeTest) {
+  // Prepare single calculator graph to and wait for packets.
+  CalculatorGraphConfig graph_config = ParseTextProtoOrDie<
+      CalculatorGraphConfig>(
+      R"(
+        node {
+          calculator: "ConstantSidePacketCalculator"
+          output_side_packet: "PACKET:model_path"
+          options: {
+            [mediapipe.ConstantSidePacketCalculatorOptions.ext]: {
+              packet {
+                string_value: "mediapipe/calculators/tflite/testdata/add.bin"
+              }
+            }
+          }
+        }
+
+        node {
+          calculator: "LocalFileContentsCalculator"
+          input_side_packet: "FILE_PATH:model_path"
+          output_side_packet: "CONTENTS:model_blob"
+        }
+
+        node {
+          calculator: "TfLiteModelCalculator"
+          input_side_packet: "MODEL_BLOB:model_blob"
+          output_side_packet: "MODEL:model"
+        }
+      )");
+  CalculatorGraph graph(graph_config);
+  MP_ASSERT_OK(graph.StartRun({}));
+  MP_ASSERT_OK(graph.WaitUntilIdle());
+  auto status_or_packet = graph.GetOutputSidePacket("model");
+  MP_ASSERT_OK(status_or_packet);
+  auto model_packet = status_or_packet.ValueOrDie();
+  const auto& model = model_packet.Get<
+      std::unique_ptr<tflite::FlatBufferModel,
+                      std::function<void(tflite::FlatBufferModel*)>>>();
+
+  auto expected_model = tflite::FlatBufferModel::BuildFromFile(
+      "mediapipe/calculators/tflite/testdata/add.bin");
+
+  EXPECT_EQ(model->GetModel()->version(),
+            expected_model->GetModel()->version());
+  EXPECT_EQ(model->GetModel()->buffers()->size(),
+            expected_model->GetModel()->buffers()->size());
+  const int num_subgraphs = expected_model->GetModel()->subgraphs()->size();
+  EXPECT_EQ(model->GetModel()->subgraphs()->size(), num_subgraphs);
+  for (int i = 0; i < num_subgraphs; ++i) {
+    const auto* expected_subgraph =
+        expected_model->GetModel()->subgraphs()->Get(i);
+    const auto* subgraph = model->GetModel()->subgraphs()->Get(i);
+    const int num_tensors = expected_subgraph->tensors()->size();
+    EXPECT_EQ(subgraph->tensors()->size(), num_tensors);
+    for (int j = 0; j < num_tensors; ++j) {
+      EXPECT_EQ(subgraph->tensors()->Get(j)->name()->str(),
+                expected_subgraph->tensors()->Get(j)->name()->str());
+    }
+  }
+}
+
+}  // namespace mediapipe

mediapipe/calculators/tflite/tflite_tensors_to_classification_calculator.cc

@@ -129,22 +129,43 @@ REGISTER_CALCULATOR(TfLiteTensorsToClassificationCalculator);
     num_classes *= raw_score_tensor->dims->data[i];
   }
 
+  if (options_.binary_classification()) {
+    RET_CHECK_EQ(num_classes, 1);
+    // Number of classes for binary classification.
+    num_classes = 2;
+  }
   if (label_map_loaded_) {
     RET_CHECK_EQ(num_classes, label_map_.size());
   }
   const float* raw_scores = raw_score_tensor->data.f;
 
   auto classification_list = absl::make_unique<ClassificationList>();
-  for (int i = 0; i < num_classes; ++i) {
-    if (options_.has_min_score_threshold() &&
-        raw_scores[i] < options_.min_score_threshold()) {
-      continue;
-    }
-    Classification* classification = classification_list->add_classification();
-    classification->set_index(i);
-    classification->set_score(raw_scores[i]);
+  if (options_.binary_classification()) {
+    Classification* class_first = classification_list->add_classification();
+    Classification* class_second = classification_list->add_classification();
+    class_first->set_index(0);
+    class_second->set_index(1);
+    class_first->set_score(raw_scores[0]);
+    class_second->set_score(1. - raw_scores[0]);
+
     if (label_map_loaded_) {
-      classification->set_label(label_map_[i]);
+      class_first->set_label(label_map_[0]);
+      class_second->set_label(label_map_[1]);
+    }
+  } else {
+    for (int i = 0; i < num_classes; ++i) {
+      if (options_.has_min_score_threshold() &&
+          raw_scores[i] < options_.min_score_threshold()) {
+        continue;
+      }
+      Classification* classification =
+          classification_list->add_classification();
+      classification->set_index(i);
+      classification->set_score(raw_scores[i]);
+
+      if (label_map_loaded_) {
+        classification->set_label(label_map_[i]);
+      }
     }
   }
 

mediapipe/calculators/tflite/tflite_tensors_to_classification_calculator.proto

@@ -32,4 +32,10 @@ message TfLiteTensorsToClassificationCalculatorOptions {
   optional int32 top_k = 2;
   // Path to a label map file for getting the actual name of class ids.
   optional string label_map_path = 3;
+  // Whether the input is a single float for binary classification.
+  // When true, only a single float is expected in the input tensor and the
+  // label map, if provided, is expected to have exactly two labels.
+  // The single score(float) represent the probability of first label, and
+  // 1 - score is the probabilility of the second label.
+  optional bool binary_classification = 4;
 }

mediapipe/calculators/tflite/tflite_tensors_to_landmarks_calculator.cc

@@ -28,6 +28,21 @@ namespace mediapipe {
 //  TENSORS - Vector of TfLiteTensor of type kTfLiteFloat32. Only the first
 //            tensor will be used. The size of the values must be
 //            (num_dimension x num_landmarks).
+//
+//  FLIP_HORIZONTALLY (optional): Whether to flip landmarks horizontally or
+//  not. Overrides corresponding side packet and/or field in the calculator
+//  options.
+//
+//  FLIP_VERTICALLY (optional): Whether to flip landmarks vertically or not.
+//  Overrides corresponding side packet and/or field in the calculator options.
+//
+// Input side packet:
+//   FLIP_HORIZONTALLY (optional): Whether to flip landmarks horizontally or
+//   not. Overrides the corresponding field in the calculator options.
+//
+//   FLIP_VERTICALLY (optional): Whether to flip landmarks vertically or not.
+//   Overrides the corresponding field in the calculator options.
+//
 // Output:
 //  LANDMARKS(optional) - Result MediaPipe landmarks.
 //  NORM_LANDMARKS(optional) - Result MediaPipe normalized landmarks.
@@ -61,6 +76,8 @@ class TfLiteTensorsToLandmarksCalculator : public CalculatorBase {
  private:
   ::mediapipe::Status LoadOptions(CalculatorContext* cc);
   int num_landmarks_ = 0;
+  bool flip_vertically_ = false;
+  bool flip_horizontally_ = false;
 
   ::mediapipe::TfLiteTensorsToLandmarksCalculatorOptions options_;
 };
@@ -75,6 +92,22 @@ REGISTER_CALCULATOR(TfLiteTensorsToLandmarksCalculator);
     cc->Inputs().Tag("TENSORS").Set<std::vector<TfLiteTensor>>();
   }
 
+  if (cc->Inputs().HasTag("FLIP_HORIZONTALLY")) {
+    cc->Inputs().Tag("FLIP_HORIZONTALLY").Set<bool>();
+  }
+
+  if (cc->Inputs().HasTag("FLIP_VERTICALLY")) {
+    cc->Inputs().Tag("FLIP_VERTICALLY").Set<bool>();
+  }
+
+  if (cc->InputSidePackets().HasTag("FLIP_HORIZONTALLY")) {
+    cc->InputSidePackets().Tag("FLIP_HORIZONTALLY").Set<bool>();
+  }
+
+  if (cc->InputSidePackets().HasTag("FLIP_VERTICALLY")) {
+    cc->InputSidePackets().Tag("FLIP_VERTICALLY").Set<bool>();
+  }
+
   if (cc->Outputs().HasTag("LANDMARKS")) {
     cc->Outputs().Tag("LANDMARKS").Set<LandmarkList>();
   }
@@ -98,17 +131,40 @@ REGISTER_CALCULATOR(TfLiteTensorsToLandmarksCalculator);
         << "Must provide input with/height for getting normalized landmarks.";
   }
   if (cc->Outputs().HasTag("LANDMARKS") &&
-      (options_.flip_vertically() || options_.flip_horizontally())) {
+      (options_.flip_vertically() || options_.flip_horizontally() ||
+       cc->InputSidePackets().HasTag("FLIP_HORIZONTALLY") ||
+       cc->InputSidePackets().HasTag("FLIP_VERTICALLY"))) {
     RET_CHECK(options_.has_input_image_height() &&
               options_.has_input_image_width())
         << "Must provide input with/height for using flip_vertically option "
            "when outputing landmarks in absolute coordinates.";
   }
+
+  flip_horizontally_ =
+      cc->InputSidePackets().HasTag("FLIP_HORIZONTALLY")
+          ? cc->InputSidePackets().Tag("FLIP_HORIZONTALLY").Get<bool>()
+          : options_.flip_horizontally();
+
+  flip_vertically_ =
+      cc->InputSidePackets().HasTag("FLIP_VERTICALLY")
+          ? cc->InputSidePackets().Tag("FLIP_VERTICALLY").Get<bool>()
+          : options_.flip_vertically();
+
   return ::mediapipe::OkStatus();
 }
 
 ::mediapipe::Status TfLiteTensorsToLandmarksCalculator::Process(
     CalculatorContext* cc) {
+  // Override values if specified so.
+  if (cc->Inputs().HasTag("FLIP_HORIZONTALLY") &&
+      !cc->Inputs().Tag("FLIP_HORIZONTALLY").IsEmpty()) {
+    flip_horizontally_ = cc->Inputs().Tag("FLIP_HORIZONTALLY").Get<bool>();
+  }
+  if (cc->Inputs().HasTag("FLIP_VERTICALLY") &&
+      !cc->Inputs().Tag("FLIP_VERTICALLY").IsEmpty()) {
+    flip_vertically_ = cc->Inputs().Tag("FLIP_VERTICALLY").Get<bool>();
+  }
+
   if (cc->Inputs().Tag("TENSORS").IsEmpty()) {
     return ::mediapipe::OkStatus();
   }
@@ -133,13 +189,13 @@ REGISTER_CALCULATOR(TfLiteTensorsToLandmarksCalculator);
     const int offset = ld * num_dimensions;
     Landmark* landmark = output_landmarks.add_landmark();
 
-    if (options_.flip_horizontally()) {
+    if (flip_horizontally_) {
       landmark->set_x(options_.input_image_width() - raw_landmarks[offset]);
     } else {
       landmark->set_x(raw_landmarks[offset]);
     }
     if (num_dimensions > 1) {
-      if (options_.flip_vertically()) {
+      if (flip_vertically_) {
         landmark->set_y(options_.input_image_height() -
                         raw_landmarks[offset + 1]);
       } else {

mediapipe/calculators/tflite/util.h

@@ -15,11 +15,11 @@
 #ifndef MEDIAPIPE_CALCULATORS_TFLITE_UTIL_H_
 #define MEDIAPIPE_CALCULATORS_TFLITE_UTIL_H_
 
-#define RET_CHECK_CALL(call)                                     \
-  do {                                                           \
-    const auto status = (call);                                  \
-    if (ABSL_PREDICT_FALSE(!status.ok()))                        \
-      return ::mediapipe::InternalError(status.error_message()); \
+#define RET_CHECK_CALL(call)                               \
+  do {                                                     \
+    const auto status = (call);                            \
+    if (ABSL_PREDICT_FALSE(!status.ok()))                  \
+      return ::mediapipe::InternalError(status.message()); \
   } while (0);
 
 #endif  // MEDIAPIPE_CALCULATORS_TFLITE_UTIL_H_

mediapipe/calculators/util/BUILD

@@ -39,6 +39,15 @@ proto_library(
     ],
 )
 
+proto_library(
+    name = "timed_box_list_id_to_label_calculator_proto",
+    srcs = ["timed_box_list_id_to_label_calculator.proto"],
+    visibility = ["//visibility:public"],
+    deps = [
+        "//mediapipe/framework:calculator_proto",
+    ],
+)
+
 proto_library(
     name = "latency_proto",
     srcs = ["latency.proto"],
@@ -113,6 +122,18 @@ mediapipe_cc_proto_library(
     ],
 )
 
+mediapipe_cc_proto_library(
+    name = "timed_box_list_id_to_label_calculator_cc_proto",
+    srcs = ["timed_box_list_id_to_label_calculator.proto"],
+    cc_deps = [
+        "//mediapipe/framework:calculator_cc_proto",
+    ],
+    visibility = ["//visibility:public"],
+    deps = [
+        ":timed_box_list_id_to_label_calculator_proto",
+    ],
+)
+
 mediapipe_cc_proto_library(
     name = "latency_cc_proto",
     srcs = ["latency.proto"],
@@ -300,7 +321,35 @@ cc_library(
         "//mediapipe:android": [
             "//mediapipe/util/android/file/base",
         ],
-        "//mediapipe:apple": [
+        "//mediapipe:ios": [
+            "//mediapipe/util/android/file/base",
+        ],
+        "//mediapipe:macos": [
+            "//mediapipe/framework/port:file_helpers",
+        ],
+        "//conditions:default": [
+            "//mediapipe/framework/port:file_helpers",
+        ],
+    }),
+    alwayslink = 1,
+)
+
+cc_library(
+    name = "timed_box_list_id_to_label_calculator",
+    srcs = ["timed_box_list_id_to_label_calculator.cc"],
+    visibility = ["//visibility:public"],
+    deps = [
+        ":timed_box_list_id_to_label_calculator_cc_proto",
+        "//mediapipe/framework/port:status",
+        "//mediapipe/framework:calculator_framework",
+        "//mediapipe/framework:packet",
+        "//mediapipe/util/tracking:box_tracker_cc_proto",
+        "//mediapipe/util:resource_util",
+    ] + select({
+        "//mediapipe:android": [
+            "//mediapipe/util/android/file/base",
+        ],
+        "//mediapipe:ios": [
             "//mediapipe/util/android/file/base",
         ],
         "//mediapipe:macos": [
@@ -437,6 +486,7 @@ cc_library(
         "//mediapipe/framework/formats:rect_cc_proto",
         "//mediapipe/framework/port:ret_check",
         "//mediapipe/framework/port:status",
+        "@com_google_absl//absl/types:optional",
     ],
     alwayslink = 1,
 )
@@ -876,7 +926,7 @@ cc_library(
         "//mediapipe:android": [
             "//mediapipe/util/android/file/base",
         ],
-        "//mediapipe:apple": [
+        "//mediapipe:ios": [
             "//mediapipe/util/android/file/base",
         ],
         "//mediapipe:macos": [
@@ -919,6 +969,19 @@ cc_library(
     name = "local_file_contents_calculator",
     srcs = ["local_file_contents_calculator.cc"],
     visibility = ["//visibility:public"],
+    deps = [
+        "//mediapipe/framework:calculator_framework",
+        "//mediapipe/framework/port:ret_check",
+        "//mediapipe/framework/port:status",
+        "//mediapipe/util:resource_util",
+    ],
+    alwayslink = 1,
+)
+
+cc_library(
+    name = "local_file_pattern_contents_calculator",
+    srcs = ["local_file_pattern_contents_calculator.cc"],
+    visibility = ["//visibility:public"],
     deps = [
         "//mediapipe/framework:calculator_framework",
         "//mediapipe/framework/port:file_helpers",
@@ -935,6 +998,7 @@ cc_library(
     visibility = ["//visibility:public"],
     deps = [
         "//mediapipe/framework:calculator_framework",
+        "//mediapipe/framework/formats:classification_cc_proto",
         "//mediapipe/framework/formats:landmark_cc_proto",
         "//mediapipe/framework/formats:rect_cc_proto",
         "//mediapipe/framework/port:ret_check",
@@ -952,6 +1016,7 @@ cc_library(
     deps = [
         ":collection_has_min_size_calculator_cc_proto",
         "//mediapipe/framework:calculator_framework",
+        "//mediapipe/framework/formats:landmark_cc_proto",
         "//mediapipe/framework/formats:rect_cc_proto",
         "//mediapipe/framework/port:ret_check",
         "//mediapipe/framework/port:status",
@@ -959,6 +1024,18 @@ cc_library(
     alwayslink = 1,
 )
 
+cc_test(
+    name = "collection_has_min_size_calculator_test",
+    srcs = ["collection_has_min_size_calculator_test.cc"],
+    deps = [
+        ":collection_has_min_size_calculator",
+        "//mediapipe/framework:calculator_framework",
+        "//mediapipe/framework:calculator_runner",
+        "//mediapipe/framework/port:gtest_main",
+        "//mediapipe/framework/port:parse_text_proto",
+    ],
+)
+
 cc_library(
     name = "association_calculator",
     hdrs = ["association_calculator.h"],

mediapipe/calculators/util/annotation_overlay_calculator.cc

@@ -39,13 +39,13 @@ namespace mediapipe {
 
 namespace {
 
-constexpr char kInputFrameTag[] = "INPUT_FRAME";
-constexpr char kOutputFrameTag[] = "OUTPUT_FRAME";
+constexpr char kInputFrameTag[] = "IMAGE";
+constexpr char kOutputFrameTag[] = "IMAGE";
 
 constexpr char kInputVectorTag[] = "VECTOR";
 
-constexpr char kInputFrameTagGpu[] = "INPUT_FRAME_GPU";
-constexpr char kOutputFrameTagGpu[] = "OUTPUT_FRAME_GPU";
+constexpr char kInputFrameTagGpu[] = "IMAGE_GPU";
+constexpr char kOutputFrameTagGpu[] = "IMAGE_GPU";
 
 enum { ATTRIB_VERTEX, ATTRIB_TEXTURE_POSITION, NUM_ATTRIBUTES };
 
@@ -55,13 +55,13 @@ size_t RoundUp(size_t n, size_t m) { return ((n + m - 1) / m) * m; }  // NOLINT
 // When using GPU, this color will become transparent when the calculator
 // merges the annotation overlay with the image frame. As a result, drawing in
 // this color is not supported and it should be set to something unlikely used.
-constexpr int kAnnotationBackgroundColor[] = {100, 101, 102};
+constexpr uchar kAnnotationBackgroundColor = 2;  // Grayscale value.
 }  // namespace
 
 // A calculator for rendering data on images.
 //
 // Inputs:
-//  1. INPUT_FRAME or INPUT_FRAME_GPU (optional): An ImageFrame (or GpuBuffer)
+//  1. IMAGE or IMAGE_GPU (optional): An ImageFrame (or GpuBuffer)
 //     containing the input image.
 //     If output is CPU, and input isn't provided, the renderer creates a
 //     blank canvas with the width, height and color provided in the options.
@@ -73,7 +73,7 @@ constexpr int kAnnotationBackgroundColor[] = {100, 101, 102};
 //     input vector items. These input streams are tagged with "VECTOR".
 //
 // Output:
-//  1. OUTPUT_FRAME or OUTPUT_FRAME_GPU: A rendered ImageFrame (or GpuBuffer).
+//  1. IMAGE or IMAGE_GPU: A rendered ImageFrame (or GpuBuffer).
 //
 // For CPU input frames, only SRGBA, SRGB and GRAY8 format are supported. The
 // output format is the same as input except for GRAY8 where the output is in
@@ -87,13 +87,13 @@ constexpr int kAnnotationBackgroundColor[] = {100, 101, 102};
 // Example config (CPU):
 // node {
 //   calculator: "AnnotationOverlayCalculator"
-//   input_stream: "INPUT_FRAME:image_frames"
+//   input_stream: "IMAGE:image_frames"
 //   input_stream: "render_data_1"
 //   input_stream: "render_data_2"
 //   input_stream: "render_data_3"
 //   input_stream: "VECTOR:0:render_data_vec_0"
 //   input_stream: "VECTOR:1:render_data_vec_1"
-//   output_stream: "OUTPUT_FRAME:decorated_frames"
+//   output_stream: "IMAGE:decorated_frames"
 //   options {
 //     [mediapipe.AnnotationOverlayCalculatorOptions.ext] {
 //     }
@@ -103,13 +103,13 @@ constexpr int kAnnotationBackgroundColor[] = {100, 101, 102};
 // Example config (GPU):
 // node {
 //   calculator: "AnnotationOverlayCalculator"
-//   input_stream: "INPUT_FRAME_GPU:image_frames"
+//   input_stream: "IMAGE_GPU:image_frames"
 //   input_stream: "render_data_1"
 //   input_stream: "render_data_2"
 //   input_stream: "render_data_3"
 //   input_stream: "VECTOR:0:render_data_vec_0"
 //   input_stream: "VECTOR:1:render_data_vec_1"
-//   output_stream: "OUTPUT_FRAME_GPU:decorated_frames"
+//   output_stream: "IMAGE_GPU:decorated_frames"
 //   options {
 //     [mediapipe.AnnotationOverlayCalculatorOptions.ext] {
 //     }
@@ -491,11 +491,9 @@ REGISTER_CALCULATOR(AnnotationOverlayCalculator);
     if (format != mediapipe::ImageFormat::SRGBA &&
         format != mediapipe::ImageFormat::SRGB)
       RET_CHECK_FAIL() << "Unsupported GPU input format: " << format;
-
-    image_mat = absl::make_unique<cv::Mat>(
-        height_, width_, CV_8UC3,
-        cv::Scalar(kAnnotationBackgroundColor[0], kAnnotationBackgroundColor[1],
-                   kAnnotationBackgroundColor[2]));
+    image_mat = absl::make_unique<cv::Mat>(height_, width_, CV_8UC3);
+    memset(image_mat->data, kAnnotationBackgroundColor,
+           height_ * width_ * image_mat->elemSize());
   } else {
     image_mat = absl::make_unique<cv::Mat>(
         options_.canvas_height_px(), options_.canvas_width_px(), CV_8UC3,
@@ -617,9 +615,9 @@ REGISTER_CALCULATOR(AnnotationOverlayCalculator);
   glUniform1i(glGetUniformLocation(program_, "input_frame"), 1);
   glUniform1i(glGetUniformLocation(program_, "overlay"), 2);
   glUniform3f(glGetUniformLocation(program_, "transparent_color"),
-              kAnnotationBackgroundColor[0] / 255.0,
-              kAnnotationBackgroundColor[1] / 255.0,
-              kAnnotationBackgroundColor[2] / 255.0);
+              kAnnotationBackgroundColor / 255.0,
+              kAnnotationBackgroundColor / 255.0,
+              kAnnotationBackgroundColor / 255.0);
 
   // Init texture for opencv rendered frame.
   const auto& input_frame =

mediapipe/calculators/util/collection_has_min_size_calculator.cc

@@ -15,6 +15,9 @@
 
 #include "mediapipe/calculators/util/collection_has_min_size_calculator.h"
 
+#include <vector>
+
+#include "mediapipe/framework/formats/landmark.pb.h"
 #include "mediapipe/framework/formats/rect.pb.h"
 
 namespace mediapipe {
@@ -23,4 +26,9 @@ typedef CollectionHasMinSizeCalculator<std::vector<::mediapipe::NormalizedRect>>
     NormalizedRectVectorHasMinSizeCalculator;
 REGISTER_CALCULATOR(NormalizedRectVectorHasMinSizeCalculator);
 
+typedef CollectionHasMinSizeCalculator<
+    std::vector<::mediapipe::NormalizedLandmarkList>>
+    NormalizedLandmarkListVectorHasMinSizeCalculator;
+REGISTER_CALCULATOR(NormalizedLandmarkListVectorHasMinSizeCalculator);
+
 }  // namespace mediapipe

mediapipe/calculators/util/collection_has_min_size_calculator_test.cc

@@ -0,0 +1,156 @@
+// Copyright 2020 The MediaPipe Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "mediapipe/calculators/util/collection_has_min_size_calculator.h"
+
+#include <memory>
+#include <string>
+#include <vector>
+
+#include "mediapipe/framework/calculator_framework.h"
+#include "mediapipe/framework/calculator_runner.h"
+#include "mediapipe/framework/port/gmock.h"
+#include "mediapipe/framework/port/gtest.h"
+#include "mediapipe/framework/port/parse_text_proto.h"
+#include "mediapipe/framework/port/status_matchers.h"  // NOLINT
+
+namespace mediapipe {
+
+typedef CollectionHasMinSizeCalculator<std::vector<int>>
+    TestIntCollectionHasMinSizeCalculator;
+REGISTER_CALCULATOR(TestIntCollectionHasMinSizeCalculator);
+
+void AddInputVector(const std::vector<int>& input, int64 timestamp,
+                    CalculatorRunner* runner) {
+  runner->MutableInputs()
+      ->Tag("ITERABLE")
+      .packets.push_back(
+          MakePacket<std::vector<int>>(input).At(Timestamp(timestamp)));
+}
+
+TEST(TestIntCollectionHasMinSizeCalculator, DoesHaveMinSize) {
+  CalculatorGraphConfig::Node node_config =
+      ParseTextProtoOrDie<CalculatorGraphConfig::Node>(R"(
+        calculator: "TestIntCollectionHasMinSizeCalculator"
+        input_stream: "ITERABLE:input_vector"
+        output_stream: "output_vector"
+        options {
+          [mediapipe.CollectionHasMinSizeCalculatorOptions.ext] { min_size: 2 }
+        }
+      )");
+  CalculatorRunner runner(node_config);
+  const std::vector<Packet>& outputs = runner.Outputs().Index(0).packets;
+
+  AddInputVector({1, 2}, /*timestamp=*/1, &runner);
+  MP_ASSERT_OK(runner.Run());
+
+  EXPECT_EQ(1, outputs.size());
+  EXPECT_EQ(Timestamp(1), outputs[0].Timestamp());
+  EXPECT_TRUE(outputs[0].Get<bool>());
+
+  AddInputVector({1, 2, 3}, /*timestamp=*/2, &runner);
+  MP_ASSERT_OK(runner.Run());
+
+  EXPECT_EQ(2, outputs.size());
+  EXPECT_EQ(Timestamp(2), outputs[1].Timestamp());
+  EXPECT_TRUE(outputs[1].Get<bool>());
+}
+
+TEST(TestIntCollectionHasMinSizeCalculator,
+     DoesHaveMinSize_MinSizeAsSidePacket) {
+  CalculatorGraphConfig::Node node_config =
+      ParseTextProtoOrDie<CalculatorGraphConfig::Node>(R"(
+        calculator: "TestIntCollectionHasMinSizeCalculator"
+        input_stream: "ITERABLE:input_vector"
+        input_side_packet: "min_size"
+        output_stream: "output_vector"
+      )");
+  CalculatorRunner runner(node_config);
+  const std::vector<Packet>& outputs = runner.Outputs().Index(0).packets;
+
+  runner.MutableSidePackets()->Index(0) = MakePacket<int>(2);
+
+  AddInputVector({1, 2}, /*timestamp=*/1, &runner);
+  MP_ASSERT_OK(runner.Run());
+
+  EXPECT_EQ(1, outputs.size());
+  EXPECT_EQ(Timestamp(1), outputs[0].Timestamp());
+  EXPECT_TRUE(outputs[0].Get<bool>());
+
+  AddInputVector({1, 2, 3}, /*timestamp=*/2, &runner);
+  MP_ASSERT_OK(runner.Run());
+
+  EXPECT_EQ(2, outputs.size());
+  EXPECT_EQ(Timestamp(2), outputs[1].Timestamp());
+  EXPECT_TRUE(outputs[1].Get<bool>());
+}
+
+TEST(TestIntCollectionHasMinSizeCalculator, DoesNotHaveMinSize) {
+  CalculatorGraphConfig::Node node_config =
+      ParseTextProtoOrDie<CalculatorGraphConfig::Node>(R"(
+        calculator: "TestIntCollectionHasMinSizeCalculator"
+        input_stream: "ITERABLE:input_vector"
+        output_stream: "output_vector"
+        options {
+          [mediapipe.CollectionHasMinSizeCalculatorOptions.ext] { min_size: 3 }
+        }
+      )");
+  CalculatorRunner runner(node_config);
+  const std::vector<Packet>& outputs = runner.Outputs().Index(0).packets;
+
+  AddInputVector({1}, /*timestamp=*/1, &runner);
+  MP_ASSERT_OK(runner.Run());
+
+  EXPECT_EQ(1, outputs.size());
+  EXPECT_EQ(Timestamp(1), outputs[0].Timestamp());
+  EXPECT_FALSE(outputs[0].Get<bool>());
+
+  AddInputVector({1, 2}, /*timestamp=*/2, &runner);
+  MP_ASSERT_OK(runner.Run());
+
+  EXPECT_EQ(2, outputs.size());
+  EXPECT_EQ(Timestamp(2), outputs[1].Timestamp());
+  EXPECT_FALSE(outputs[1].Get<bool>());
+}
+
+TEST(TestIntCollectionHasMinSizeCalculator,
+     DoesNotHaveMinSize_MinSizeAsSidePacket) {
+  CalculatorGraphConfig::Node node_config =
+      ParseTextProtoOrDie<CalculatorGraphConfig::Node>(R"(
+        calculator: "TestIntCollectionHasMinSizeCalculator"
+        input_stream: "ITERABLE:input_vector"
+        input_side_packet: "min_size"
+        output_stream: "output_vector"
+      )");
+  CalculatorRunner runner(node_config);
+  const std::vector<Packet>& outputs = runner.Outputs().Index(0).packets;
+
+  runner.MutableSidePackets()->Index(0) = MakePacket<int>(3);
+
+  AddInputVector({1}, /*timestamp=*/1, &runner);
+  MP_ASSERT_OK(runner.Run());
+
+  EXPECT_EQ(1, outputs.size());
+  EXPECT_EQ(Timestamp(1), outputs[0].Timestamp());
+  EXPECT_FALSE(outputs[0].Get<bool>());
+
+  AddInputVector({1, 2}, /*timestamp=*/2, &runner);
+  MP_ASSERT_OK(runner.Run());
+
+  EXPECT_EQ(2, outputs.size());
+  EXPECT_EQ(Timestamp(2), outputs[1].Timestamp());
+  EXPECT_FALSE(outputs[1].Get<bool>());
+}
+
+}  // namespace mediapipe

mediapipe/calculators/util/detection_label_id_to_text_calculator.cc

@@ -12,10 +12,10 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "mediapipe//framework/packet.h"
 #include "mediapipe/calculators/util/detection_label_id_to_text_calculator.pb.h"
 #include "mediapipe/framework/calculator_framework.h"
 #include "mediapipe/framework/formats/detection.pb.h"
+#include "mediapipe/framework/packet.h"
 #include "mediapipe/framework/port/status.h"
 #include "mediapipe/util/resource_util.h"
 

mediapipe/calculators/util/detections_to_rects_calculator.cc

@@ -39,7 +39,8 @@ constexpr char kNormRectsTag[] = "NORM_RECTS";
 }  // namespace
 
 ::mediapipe::Status DetectionsToRectsCalculator::DetectionToRect(
-    const Detection& detection, Rect* rect) {
+    const Detection& detection, const DetectionSpec& detection_spec,
+    Rect* rect) {
   const LocationData location_data = detection.location_data();
   RET_CHECK(location_data.format() == LocationData::BOUNDING_BOX)
       << "Only Detection with formats of BOUNDING_BOX can be converted to Rect";
@@ -52,7 +53,8 @@ constexpr char kNormRectsTag[] = "NORM_RECTS";
 }
 
 ::mediapipe::Status DetectionsToRectsCalculator::DetectionToNormalizedRect(
-    const Detection& detection, NormalizedRect* rect) {
+    const Detection& detection, const DetectionSpec& detection_spec,
+    NormalizedRect* rect) {
   const LocationData location_data = detection.location_data();
   RET_CHECK(location_data.format() == LocationData::RELATIVE_BOUNDING_BOX)
       << "Only Detection with formats of RELATIVE_BOUNDING_BOX can be "
@@ -174,27 +176,31 @@ constexpr char kNormRectsTag[] = "NORM_RECTS";
     }
   }
 
-  std::pair<int, int> image_size;
-  if (rotate_) {
-    RET_CHECK(!cc->Inputs().Tag(kImageSizeTag).IsEmpty());
-    image_size = cc->Inputs().Tag(kImageSizeTag).Get<std::pair<int, int>>();
-  }
+  // Get dynamic calculator options (e.g. `image_size`).
+  const DetectionSpec detection_spec = GetDetectionSpec(cc);
 
   if (cc->Outputs().HasTag(kRectTag)) {
     auto output_rect = absl::make_unique<Rect>();
-    MP_RETURN_IF_ERROR(DetectionToRect(detections[0], output_rect.get()));
+    MP_RETURN_IF_ERROR(
+        DetectionToRect(detections[0], detection_spec, output_rect.get()));
     if (rotate_) {
-      output_rect->set_rotation(ComputeRotation(detections[0], image_size));
+      float rotation;
+      MP_RETURN_IF_ERROR(
+          ComputeRotation(detections[0], detection_spec, &rotation));
+      output_rect->set_rotation(rotation);
     }
     cc->Outputs().Tag(kRectTag).Add(output_rect.release(),
                                     cc->InputTimestamp());
   }
   if (cc->Outputs().HasTag(kNormRectTag)) {
     auto output_rect = absl::make_unique<NormalizedRect>();
-    MP_RETURN_IF_ERROR(
-        DetectionToNormalizedRect(detections[0], output_rect.get()));
+    MP_RETURN_IF_ERROR(DetectionToNormalizedRect(detections[0], detection_spec,
+                                                 output_rect.get()));
     if (rotate_) {
-      output_rect->set_rotation(ComputeRotation(detections[0], image_size));
+      float rotation;
+      MP_RETURN_IF_ERROR(
+          ComputeRotation(detections[0], detection_spec, &rotation));
+      output_rect->set_rotation(rotation);
     }
     cc->Outputs()
         .Tag(kNormRectTag)
@@ -203,11 +209,13 @@ constexpr char kNormRectsTag[] = "NORM_RECTS";
   if (cc->Outputs().HasTag(kRectsTag)) {
     auto output_rects = absl::make_unique<std::vector<Rect>>(detections.size());
     for (int i = 0; i < detections.size(); ++i) {
-      MP_RETURN_IF_ERROR(
-          DetectionToRect(detections[i], &(output_rects->at(i))));
+      MP_RETURN_IF_ERROR(DetectionToRect(detections[i], detection_spec,
+                                         &(output_rects->at(i))));
       if (rotate_) {
-        output_rects->at(i).set_rotation(
-            ComputeRotation(detections[i], image_size));
+        float rotation;
+        MP_RETURN_IF_ERROR(
+            ComputeRotation(detections[i], detection_spec, &rotation));
+        output_rects->at(i).set_rotation(rotation);
       }
     }
     cc->Outputs().Tag(kRectsTag).Add(output_rects.release(),
@@ -217,11 +225,13 @@ constexpr char kNormRectsTag[] = "NORM_RECTS";
     auto output_rects =
         absl::make_unique<std::vector<NormalizedRect>>(detections.size());
     for (int i = 0; i < detections.size(); ++i) {
-      MP_RETURN_IF_ERROR(
-          DetectionToNormalizedRect(detections[i], &(output_rects->at(i))));
+      MP_RETURN_IF_ERROR(DetectionToNormalizedRect(
+          detections[i], detection_spec, &(output_rects->at(i))));
       if (rotate_) {
-        output_rects->at(i).set_rotation(
-            ComputeRotation(detections[i], image_size));
+        float rotation;
+        MP_RETURN_IF_ERROR(
+            ComputeRotation(detections[i], detection_spec, &rotation));
+        output_rects->at(i).set_rotation(rotation);
       }
     }
     cc->Outputs()
@@ -232,21 +242,35 @@ constexpr char kNormRectsTag[] = "NORM_RECTS";
   return ::mediapipe::OkStatus();
 }
 
-float DetectionsToRectsCalculator::ComputeRotation(
-    const Detection& detection, const std::pair<int, int> image_size) {
+::mediapipe::Status DetectionsToRectsCalculator::ComputeRotation(
+    const Detection& detection, const DetectionSpec& detection_spec,
+    float* rotation) {
   const auto& location_data = detection.location_data();
+  const auto& image_size = detection_spec.image_size;
+  RET_CHECK(image_size) << "Image size is required to calculate rotation";
+
   const float x0 = location_data.relative_keypoints(start_keypoint_index_).x() *
-                   image_size.first;
+                   image_size->first;
   const float y0 = location_data.relative_keypoints(start_keypoint_index_).y() *
-                   image_size.second;
+                   image_size->second;
   const float x1 = location_data.relative_keypoints(end_keypoint_index_).x() *
-                   image_size.first;
+                   image_size->first;
   const float y1 = location_data.relative_keypoints(end_keypoint_index_).y() *
-                   image_size.second;
+                   image_size->second;
 
-  float rotation = target_angle_ - std::atan2(-(y1 - y0), x1 - x0);
+  *rotation = NormalizeRadians(target_angle_ - std::atan2(-(y1 - y0), x1 - x0));
 
-  return NormalizeRadians(rotation);
+  return ::mediapipe::OkStatus();
+}
+
+DetectionSpec DetectionsToRectsCalculator::GetDetectionSpec(
+    const CalculatorContext* cc) {
+  absl::optional<std::pair<int, int>> image_size;
+  if (cc->Inputs().HasTag(kImageSizeTag)) {
+    image_size = cc->Inputs().Tag(kImageSizeTag).Get<std::pair<int, int>>();
+  }
+
+  return {image_size};
 }
 
 REGISTER_CALCULATOR(DetectionsToRectsCalculator);

mediapipe/calculators/util/detections_to_rects_calculator.h

@@ -16,6 +16,7 @@
 
 #include <cmath>
 
+#include "absl/types/optional.h"
 #include "mediapipe/calculators/util/detections_to_rects_calculator.pb.h"
 #include "mediapipe/framework/calculator_framework.h"
 #include "mediapipe/framework/calculator_options.pb.h"
@@ -27,6 +28,13 @@
 
 namespace mediapipe {
 
+// Dynamic options passed as calculator `input_stream` that can be used for
+// calculation of rectangle or rotation for given detection. Does not include
+// static calculator options which are available via private fields.
+struct DetectionSpec {
+  absl::optional<std::pair<int, int>> image_size;
+};
+
 // A calculator that converts Detection proto to Rect proto.
 //
 // Detection is the format for encoding one or more detections in an image.
@@ -81,13 +89,16 @@ class DetectionsToRectsCalculator : public CalculatorBase {
   ::mediapipe::Status Process(CalculatorContext* cc) override;
 
  protected:
-  virtual float ComputeRotation(const ::mediapipe::Detection& detection,
-                                const std::pair<int, int> image_size);
   virtual ::mediapipe::Status DetectionToRect(
-      const ::mediapipe::Detection& detection, ::mediapipe::Rect* rect);
+      const ::mediapipe::Detection& detection,
+      const DetectionSpec& detection_spec, ::mediapipe::Rect* rect);
   virtual ::mediapipe::Status DetectionToNormalizedRect(
       const ::mediapipe::Detection& detection,
-      ::mediapipe::NormalizedRect* rect);
+      const DetectionSpec& detection_spec, ::mediapipe::NormalizedRect* rect);
+  virtual ::mediapipe::Status ComputeRotation(
+      const ::mediapipe::Detection& detection,
+      const DetectionSpec& detection_spec, float* rotation);
+  virtual DetectionSpec GetDetectionSpec(const CalculatorContext* cc);
 
   static inline float NormalizeRadians(float angle) {
     return angle - 2 * M_PI * std::floor((angle - (-M_PI)) / (2 * M_PI));

mediapipe/calculators/util/filter_collection_calculator.cc

@@ -17,6 +17,7 @@
 
 #include <vector>
 
+#include "mediapipe/framework/formats/classification.pb.h"
 #include "mediapipe/framework/formats/landmark.pb.h"
 #include "mediapipe/framework/formats/rect.pb.h"
 
@@ -31,4 +32,8 @@ typedef FilterCollectionCalculator<
     FilterLandmarkListCollectionCalculator;
 REGISTER_CALCULATOR(FilterLandmarkListCollectionCalculator);
 
+typedef FilterCollectionCalculator<std::vector<::mediapipe::ClassificationList>>
+    FilterClassificationListCollectionCalculator;
+REGISTER_CALCULATOR(FilterClassificationListCollectionCalculator);
+
 }  // namespace mediapipe

mediapipe/calculators/util/labels_to_render_data_calculator.cc

@@ -128,16 +128,19 @@ REGISTER_CALCULATOR(LabelsToRenderDataCalculator);
   } else {
     const std::vector<std::string>& label_vector =
         cc->Inputs().Tag("LABELS").Get<std::vector<std::string>>();
-    std::vector<float> score_vector;
-    if (cc->Inputs().HasTag("SCORES")) {
-      score_vector = cc->Inputs().Tag("SCORES").Get<std::vector<float>>();
-    }
-    CHECK_EQ(label_vector.size(), score_vector.size());
     labels.resize(label_vector.size());
-    scores.resize(label_vector.size());
     for (int i = 0; i < label_vector.size(); ++i) {
       labels[i] = label_vector[i];
-      scores[i] = score_vector[i];
+    }
+
+    if (cc->Inputs().HasTag("SCORES")) {
+      std::vector<float> score_vector =
+          cc->Inputs().Tag("SCORES").Get<std::vector<float>>();
+      CHECK_EQ(label_vector.size(), score_vector.size());
+      scores.resize(label_vector.size());
+      for (int i = 0; i < label_vector.size(); ++i) {
+        scores[i] = score_vector[i];
+      }
     }
   }
 

mediapipe/calculators/util/landmark_letterbox_removal_calculator.cc

@@ -12,20 +12,6 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-// Copyright 2019 The MediaPipe Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
 #include <cmath>
 #include <vector>
 
@@ -67,6 +53,15 @@ constexpr char kLetterboxPaddingTag[] = "LETTERBOX_PADDING";
 //   input_stream: "LETTERBOX_PADDING:letterbox_padding"
 //   output_stream: "LANDMARKS:adjusted_landmarks"
 // }
+//
+// node {
+//   calculator: "LandmarkLetterboxRemovalCalculator"
+//   input_stream: "LANDMARKS:0:landmarks_0"
+//   input_stream: "LANDMARKS:1:landmarks_1"
+//   input_stream: "LETTERBOX_PADDING:letterbox_padding"
+//   output_stream: "LANDMARKS:0:adjusted_landmarks_0"
+//   output_stream: "LANDMARKS:1:adjusted_landmarks_1"
+// }
 class LandmarkLetterboxRemovalCalculator : public CalculatorBase {
  public:
   static ::mediapipe::Status GetContract(CalculatorContract* cc) {
@@ -74,10 +69,20 @@ class LandmarkLetterboxRemovalCalculator : public CalculatorBase {
               cc->Inputs().HasTag(kLetterboxPaddingTag))
         << "Missing one or more input streams.";
 
-    cc->Inputs().Tag(kLandmarksTag).Set<NormalizedLandmarkList>();
+    RET_CHECK_EQ(cc->Inputs().NumEntries(kLandmarksTag),
+                 cc->Outputs().NumEntries(kLandmarksTag))
+        << "Same number of input and output landmarks is required.";
+
+    for (CollectionItemId id = cc->Inputs().BeginId(kLandmarksTag);
+         id != cc->Inputs().EndId(kLandmarksTag); ++id) {
+      cc->Inputs().Get(id).Set<NormalizedLandmarkList>();
+    }
     cc->Inputs().Tag(kLetterboxPaddingTag).Set<std::array<float, 4>>();
 
-    cc->Outputs().Tag(kLandmarksTag).Set<NormalizedLandmarkList>();
+    for (CollectionItemId id = cc->Outputs().BeginId(kLandmarksTag);
+         id != cc->Outputs().EndId(kLandmarksTag); ++id) {
+      cc->Outputs().Get(id).Set<NormalizedLandmarkList>();
+    }
 
     return ::mediapipe::OkStatus();
   }
@@ -89,38 +94,45 @@ class LandmarkLetterboxRemovalCalculator : public CalculatorBase {
   }
 
   ::mediapipe::Status Process(CalculatorContext* cc) override {
-    // Only process if there's input landmarks.
-    if (cc->Inputs().Tag(kLandmarksTag).IsEmpty()) {
+    if (cc->Inputs().Tag(kLetterboxPaddingTag).IsEmpty()) {
       return ::mediapipe::OkStatus();
     }
-
-    const NormalizedLandmarkList& input_landmarks =
-        cc->Inputs().Tag(kLandmarksTag).Get<NormalizedLandmarkList>();
     const auto& letterbox_padding =
         cc->Inputs().Tag(kLetterboxPaddingTag).Get<std::array<float, 4>>();
-
     const float left = letterbox_padding[0];
     const float top = letterbox_padding[1];
     const float left_and_right = letterbox_padding[0] + letterbox_padding[2];
     const float top_and_bottom = letterbox_padding[1] + letterbox_padding[3];
 
-    NormalizedLandmarkList output_landmarks;
-    for (int i = 0; i < input_landmarks.landmark_size(); ++i) {
-      const NormalizedLandmark& landmark = input_landmarks.landmark(i);
-      NormalizedLandmark* new_landmark = output_landmarks.add_landmark();
-      const float new_x = (landmark.x() - left) / (1.0f - left_and_right);
-      const float new_y = (landmark.y() - top) / (1.0f - top_and_bottom);
+    CollectionItemId input_id = cc->Inputs().BeginId(kLandmarksTag);
+    CollectionItemId output_id = cc->Outputs().BeginId(kLandmarksTag);
+    // Number of inputs and outpus is the same according to the contract.
+    for (; input_id != cc->Inputs().EndId(kLandmarksTag);
+         ++input_id, ++output_id) {
+      const auto& input_packet = cc->Inputs().Get(input_id);
+      if (input_packet.IsEmpty()) {
+        continue;
+      }
 
-      new_landmark->set_x(new_x);
-      new_landmark->set_y(new_y);
-      // Keep z-coord as is.
-      new_landmark->set_z(landmark.z());
+      const NormalizedLandmarkList& input_landmarks =
+          input_packet.Get<NormalizedLandmarkList>();
+      NormalizedLandmarkList output_landmarks;
+      for (int i = 0; i < input_landmarks.landmark_size(); ++i) {
+        const NormalizedLandmark& landmark = input_landmarks.landmark(i);
+        NormalizedLandmark* new_landmark = output_landmarks.add_landmark();
+        const float new_x = (landmark.x() - left) / (1.0f - left_and_right);
+        const float new_y = (landmark.y() - top) / (1.0f - top_and_bottom);
+
+        new_landmark->set_x(new_x);
+        new_landmark->set_y(new_y);
+        // Keep z-coord as is.
+        new_landmark->set_z(landmark.z());
+      }
+
+      cc->Outputs().Get(output_id).AddPacket(
+          MakePacket<NormalizedLandmarkList>(output_landmarks)
+              .At(cc->InputTimestamp()));
     }
-
-    cc->Outputs()
-        .Tag(kLandmarksTag)
-        .AddPacket(MakePacket<NormalizedLandmarkList>(output_landmarks)
-                       .At(cc->InputTimestamp()));
     return ::mediapipe::OkStatus();
   }
 };

mediapipe/calculators/util/landmark_projection_calculator.cc

@@ -12,20 +12,6 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-// Copyright 2019 The MediaPipe Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
 #include <cmath>
 #include <vector>
 
@@ -63,6 +49,15 @@ constexpr char kRectTag[] = "NORM_RECT";
 //   input_stream: "NORM_RECT:rect"
 //   output_stream: "NORM_LANDMARKS:projected_landmarks"
 // }
+//
+// node {
+//   calculator: "LandmarkProjectionCalculator"
+//   input_stream: "NORM_LANDMARKS:0:landmarks_0"
+//   input_stream: "NORM_LANDMARKS:1:landmarks_1"
+//   input_stream: "NORM_RECT:rect"
+//   output_stream: "NORM_LANDMARKS:0:projected_landmarks_0"
+//   output_stream: "NORM_LANDMARKS:1:projected_landmarks_1"
+// }
 class LandmarkProjectionCalculator : public CalculatorBase {
  public:
   static ::mediapipe::Status GetContract(CalculatorContract* cc) {
@@ -70,10 +65,20 @@ class LandmarkProjectionCalculator : public CalculatorBase {
               cc->Inputs().HasTag(kRectTag))
         << "Missing one or more input streams.";
 
-    cc->Inputs().Tag(kLandmarksTag).Set<NormalizedLandmarkList>();
+    RET_CHECK_EQ(cc->Inputs().NumEntries(kLandmarksTag),
+                 cc->Outputs().NumEntries(kLandmarksTag))
+        << "Same number of input and output landmarks is required.";
+
+    for (CollectionItemId id = cc->Inputs().BeginId(kLandmarksTag);
+         id != cc->Inputs().EndId(kLandmarksTag); ++id) {
+      cc->Inputs().Get(id).Set<NormalizedLandmarkList>();
+    }
     cc->Inputs().Tag(kRectTag).Set<NormalizedRect>();
 
-    cc->Outputs().Tag(kLandmarksTag).Set<NormalizedLandmarkList>();
+    for (CollectionItemId id = cc->Outputs().BeginId(kLandmarksTag);
+         id != cc->Outputs().EndId(kLandmarksTag); ++id) {
+      cc->Outputs().Get(id).Set<NormalizedLandmarkList>();
+    }
 
     return ::mediapipe::OkStatus();
   }
@@ -85,41 +90,50 @@ class LandmarkProjectionCalculator : public CalculatorBase {
   }
 
   ::mediapipe::Status Process(CalculatorContext* cc) override {
-    const auto& options =
-        cc->Options<::mediapipe::LandmarkProjectionCalculatorOptions>();
-    // Only process if there's input landmarks.
-    if (cc->Inputs().Tag(kLandmarksTag).IsEmpty()) {
+    if (cc->Inputs().Tag(kRectTag).IsEmpty()) {
       return ::mediapipe::OkStatus();
     }
-
-    const NormalizedLandmarkList& input_landmarks =
-        cc->Inputs().Tag(kLandmarksTag).Get<NormalizedLandmarkList>();
     const auto& input_rect = cc->Inputs().Tag(kRectTag).Get<NormalizedRect>();
 
-    NormalizedLandmarkList output_landmarks;
-    for (int i = 0; i < input_landmarks.landmark_size(); ++i) {
-      const NormalizedLandmark& landmark = input_landmarks.landmark(i);
-      NormalizedLandmark* new_landmark = output_landmarks.add_landmark();
+    const auto& options =
+        cc->Options<::mediapipe::LandmarkProjectionCalculatorOptions>();
 
-      const float x = landmark.x() - 0.5f;
-      const float y = landmark.y() - 0.5f;
-      const float angle = options.ignore_rotation() ? 0 : input_rect.rotation();
-      float new_x = std::cos(angle) * x - std::sin(angle) * y;
-      float new_y = std::sin(angle) * x + std::cos(angle) * y;
+    CollectionItemId input_id = cc->Inputs().BeginId(kLandmarksTag);
+    CollectionItemId output_id = cc->Outputs().BeginId(kLandmarksTag);
+    // Number of inputs and outpus is the same according to the contract.
+    for (; input_id != cc->Inputs().EndId(kLandmarksTag);
+         ++input_id, ++output_id) {
+      const auto& input_packet = cc->Inputs().Get(input_id);
+      if (input_packet.IsEmpty()) {
+        continue;
+      }
 
-      new_x = new_x * input_rect.width() + input_rect.x_center();
-      new_y = new_y * input_rect.height() + input_rect.y_center();
+      const auto& input_landmarks = input_packet.Get<NormalizedLandmarkList>();
+      NormalizedLandmarkList output_landmarks;
+      for (int i = 0; i < input_landmarks.landmark_size(); ++i) {
+        const NormalizedLandmark& landmark = input_landmarks.landmark(i);
+        NormalizedLandmark* new_landmark = output_landmarks.add_landmark();
 
-      new_landmark->set_x(new_x);
-      new_landmark->set_y(new_y);
-      // Keep z-coord as is.
-      new_landmark->set_z(landmark.z());
+        const float x = landmark.x() - 0.5f;
+        const float y = landmark.y() - 0.5f;
+        const float angle =
+            options.ignore_rotation() ? 0 : input_rect.rotation();
+        float new_x = std::cos(angle) * x - std::sin(angle) * y;
+        float new_y = std::sin(angle) * x + std::cos(angle) * y;
+
+        new_x = new_x * input_rect.width() + input_rect.x_center();
+        new_y = new_y * input_rect.height() + input_rect.y_center();
+
+        new_landmark->set_x(new_x);
+        new_landmark->set_y(new_y);
+        // Keep z-coord as is.
+        new_landmark->set_z(landmark.z());
+      }
+
+      cc->Outputs().Get(output_id).AddPacket(
+          MakePacket<NormalizedLandmarkList>(output_landmarks)
+              .At(cc->InputTimestamp()));
     }
-
-    cc->Outputs()
-        .Tag(kLandmarksTag)
-        .AddPacket(MakePacket<NormalizedLandmarkList>(output_landmarks)
-                       .At(cc->InputTimestamp()));
     return ::mediapipe::OkStatus();
   }
 };

mediapipe/calculators/util/landmarks_to_render_data_calculator.cc

@@ -29,6 +29,7 @@ namespace {
 
 constexpr char kLandmarksTag[] = "LANDMARKS";
 constexpr char kNormLandmarksTag[] = "NORM_LANDMARKS";
+constexpr char kRenderScaleTag[] = "RENDER_SCALE";
 constexpr char kRenderDataTag[] = "RENDER_DATA";
 constexpr char kLandmarkLabel[] = "KEYPOINT";
 constexpr int kMaxLandmarkThickness = 18;
@@ -71,6 +72,83 @@ void SetColorSizeValueFromZ(float z, float z_min, float z_max,
   render_annotation->set_thickness(thickness);
 }
 
+template <class LandmarkType>
+void AddConnectionToRenderData(const LandmarkType& start,
+                               const LandmarkType& end, int gray_val1,
+                               int gray_val2, float thickness, bool normalized,
+                               RenderData* render_data) {
+  auto* connection_annotation = render_data->add_render_annotations();
+  RenderAnnotation::GradientLine* line =
+      connection_annotation->mutable_gradient_line();
+  line->set_x_start(start.x());
+  line->set_y_start(start.y());
+  line->set_x_end(end.x());
+  line->set_y_end(end.y());
+  line->set_normalized(normalized);
+  line->mutable_color1()->set_r(gray_val1);
+  line->mutable_color1()->set_g(gray_val1);
+  line->mutable_color1()->set_b(gray_val1);
+  line->mutable_color2()->set_r(gray_val2);
+  line->mutable_color2()->set_g(gray_val2);
+  line->mutable_color2()->set_b(gray_val2);
+  connection_annotation->set_thickness(thickness);
+}
+
+template <class LandmarkListType, class LandmarkType>
+void AddConnectionsWithDepth(const LandmarkListType& landmarks,
+                             const std::vector<int>& landmark_connections,
+                             float thickness, bool normalized, float min_z,
+                             float max_z, RenderData* render_data) {
+  for (int i = 0; i < landmark_connections.size(); i += 2) {
+    const auto& ld0 = landmarks.landmark(landmark_connections[i]);
+    const auto& ld1 = landmarks.landmark(landmark_connections[i + 1]);
+    const int gray_val1 =
+        255 - static_cast<int>(Remap(ld0.z(), min_z, max_z, 255));
+    const int gray_val2 =
+        255 - static_cast<int>(Remap(ld1.z(), min_z, max_z, 255));
+    AddConnectionToRenderData<LandmarkType>(ld0, ld1, gray_val1, gray_val2,
+                                            thickness, normalized, render_data);
+  }
+}
+
+template <class LandmarkType>
+void AddConnectionToRenderData(const LandmarkType& start,
+                               const LandmarkType& end,
+                               const Color& connection_color, float thickness,
+                               bool normalized, RenderData* render_data) {
+  auto* connection_annotation = render_data->add_render_annotations();
+  RenderAnnotation::Line* line = connection_annotation->mutable_line();
+  line->set_x_start(start.x());
+  line->set_y_start(start.y());
+  line->set_x_end(end.x());
+  line->set_y_end(end.y());
+  line->set_normalized(normalized);
+  SetColor(connection_annotation, connection_color);
+  connection_annotation->set_thickness(thickness);
+}
+
+template <class LandmarkListType, class LandmarkType>
+void AddConnections(const LandmarkListType& landmarks,
+                    const std::vector<int>& landmark_connections,
+                    const Color& connection_color, float thickness,
+                    bool normalized, RenderData* render_data) {
+  for (int i = 0; i < landmark_connections.size(); i += 2) {
+    const auto& ld0 = landmarks.landmark(landmark_connections[i]);
+    const auto& ld1 = landmarks.landmark(landmark_connections[i + 1]);
+    AddConnectionToRenderData<LandmarkType>(ld0, ld1, connection_color,
+                                            thickness, normalized, render_data);
+  }
+}
+
+RenderAnnotation* AddPointRenderData(const Color& landmark_color,
+                                     float thickness, RenderData* render_data) {
+  auto* landmark_data_annotation = render_data->add_render_annotations();
+  landmark_data_annotation->set_scene_tag(kLandmarkLabel);
+  SetColor(landmark_data_annotation, landmark_color);
+  landmark_data_annotation->set_thickness(thickness);
+  return landmark_data_annotation;
+}
+
 }  // namespace
 
 // A calculator that converts Landmark proto to RenderData proto for
@@ -107,29 +185,6 @@ class LandmarksToRenderDataCalculator : public CalculatorBase {
   ::mediapipe::Status Process(CalculatorContext* cc) override;
 
  private:
-  static void AddConnectionToRenderData(
-      float start_x, float start_y, float end_x, float end_y,
-      const LandmarksToRenderDataCalculatorOptions& options, bool normalized,
-      RenderData* render_data);
-  static void SetRenderAnnotationColorThickness(
-      const LandmarksToRenderDataCalculatorOptions& options,
-      RenderAnnotation* render_annotation);
-  static RenderAnnotation* AddPointRenderData(
-      const LandmarksToRenderDataCalculatorOptions& options,
-      RenderData* render_data);
-  static void AddConnectionToRenderData(
-      float start_x, float start_y, float end_x, float end_y,
-      const LandmarksToRenderDataCalculatorOptions& options, bool normalized,
-      int gray_val1, int gray_val2, RenderData* render_data);
-
-  template <class LandmarkListType>
-  void AddConnections(const LandmarkListType& landmarks, bool normalized,
-                      RenderData* render_data);
-  template <class LandmarkListType>
-  void AddConnectionsWithDepth(const LandmarkListType& landmarks,
-                               bool normalized, float min_z, float max_z,
-                               RenderData* render_data);
-
   LandmarksToRenderDataCalculatorOptions options_;
 };
 REGISTER_CALCULATOR(LandmarksToRenderDataCalculator);
@@ -150,6 +205,9 @@ REGISTER_CALCULATOR(LandmarksToRenderDataCalculator);
   if (cc->Inputs().HasTag(kNormLandmarksTag)) {
     cc->Inputs().Tag(kNormLandmarksTag).Set<NormalizedLandmarkList>();
   }
+  if (cc->Inputs().HasTag(kRenderScaleTag)) {
+    cc->Inputs().Tag(kRenderScaleTag).Set<float>();
+  }
   cc->Outputs().Tag(kRenderDataTag).Set<RenderData>();
   return ::mediapipe::OkStatus();
 }
@@ -169,11 +227,26 @@ REGISTER_CALCULATOR(LandmarksToRenderDataCalculator);
   float z_min = 0.f;
   float z_max = 0.f;
 
+  // Apply scale to `thickness` of rendered landmarks and connections to make
+  // them bigger when object (e.g. pose, hand or face) is closer/bigger and
+  // snaller when object is further/smaller.
+  float thickness = options_.thickness();
+  if (cc->Inputs().HasTag(kRenderScaleTag)) {
+    const float render_scale = cc->Inputs().Tag(kRenderScaleTag).Get<float>();
+    thickness *= render_scale;
+  }
+
+  // Parse landmarks connections to a vector.
+  RET_CHECK_EQ(options_.landmark_connections_size() % 2, 0)
+      << "Number of entries in landmark connections must be a multiple of 2";
+  std::vector<int> landmark_connections;
+  for (int i = 0; i < options_.landmark_connections_size(); i += 1) {
+    landmark_connections.push_back(options_.landmark_connections(i));
+  }
+
   if (cc->Inputs().HasTag(kLandmarksTag)) {
     const LandmarkList& landmarks =
         cc->Inputs().Tag(kLandmarksTag).Get<LandmarkList>();
-    RET_CHECK_EQ(options_.landmark_connections_size() % 2, 0)
-        << "Number of entries in landmark connections must be a multiple of 2";
     if (visualize_depth) {
       GetMinMaxZ<LandmarkList, Landmark>(landmarks, &z_min, &z_max);
     }
@@ -181,8 +254,8 @@ REGISTER_CALCULATOR(LandmarksToRenderDataCalculator);
     visualize_depth &= ((z_max - z_min) > 1e-3);
     for (int i = 0; i < landmarks.landmark_size(); ++i) {
       const Landmark& landmark = landmarks.landmark(i);
-      auto* landmark_data_render =
-          AddPointRenderData(options_, render_data.get());
+      auto* landmark_data_render = AddPointRenderData(
+          options_.landmark_color(), thickness, render_data.get());
       if (visualize_depth) {
         SetColorSizeValueFromZ(landmark.z(), z_min, z_max,
                                landmark_data_render);
@@ -193,19 +266,19 @@ REGISTER_CALCULATOR(LandmarksToRenderDataCalculator);
       landmark_data->set_y(landmark.y());
     }
     if (visualize_depth) {
-      AddConnectionsWithDepth<LandmarkList>(landmarks, /*normalized=*/false,
-                                            z_min, z_max, render_data.get());
+      AddConnectionsWithDepth<LandmarkList, Landmark>(
+          landmarks, landmark_connections, thickness, /*normalized=*/false,
+          z_min, z_max, render_data.get());
     } else {
-      AddConnections<LandmarkList>(landmarks, /*normalized=*/false,
-                                   render_data.get());
+      AddConnections<LandmarkList, Landmark>(
+          landmarks, landmark_connections, options_.connection_color(),
+          thickness, /*normalized=*/false, render_data.get());
     }
   }
 
   if (cc->Inputs().HasTag(kNormLandmarksTag)) {
     const NormalizedLandmarkList& landmarks =
         cc->Inputs().Tag(kNormLandmarksTag).Get<NormalizedLandmarkList>();
-    RET_CHECK_EQ(options_.landmark_connections_size() % 2, 0)
-        << "Number of entries in landmark connections must be a multiple of 2";
     if (visualize_depth) {
       GetMinMaxZ<NormalizedLandmarkList, NormalizedLandmark>(landmarks, &z_min,
                                                              &z_max);
@@ -214,8 +287,8 @@ REGISTER_CALCULATOR(LandmarksToRenderDataCalculator);
     visualize_depth &= ((z_max - z_min) > 1e-3);
     for (int i = 0; i < landmarks.landmark_size(); ++i) {
       const NormalizedLandmark& landmark = landmarks.landmark(i);
-      auto* landmark_data_render =
-          AddPointRenderData(options_, render_data.get());
+      auto* landmark_data_render = AddPointRenderData(
+          options_.landmark_color(), thickness, render_data.get());
       if (visualize_depth) {
         SetColorSizeValueFromZ(landmark.z(), z_min, z_max,
                                landmark_data_render);
@@ -226,11 +299,13 @@ REGISTER_CALCULATOR(LandmarksToRenderDataCalculator);
       landmark_data->set_y(landmark.y());
     }
     if (visualize_depth) {
-      AddConnectionsWithDepth<NormalizedLandmarkList>(
-          landmarks, /*normalized=*/true, z_min, z_max, render_data.get());
+      AddConnectionsWithDepth<NormalizedLandmarkList, NormalizedLandmark>(
+          landmarks, landmark_connections, thickness, /*normalized=*/true,
+          z_min, z_max, render_data.get());
     } else {
-      AddConnections<NormalizedLandmarkList>(landmarks, /*normalized=*/true,
-                                             render_data.get());
+      AddConnections<NormalizedLandmarkList, NormalizedLandmark>(
+          landmarks, landmark_connections, options_.connection_color(),
+          thickness, /*normalized=*/true, render_data.get());
     }
   }
 
@@ -240,84 +315,4 @@ REGISTER_CALCULATOR(LandmarksToRenderDataCalculator);
   return ::mediapipe::OkStatus();
 }
 
-template <class LandmarkListType>
-void LandmarksToRenderDataCalculator::AddConnectionsWithDepth(
-    const LandmarkListType& landmarks, bool normalized, float min_z,
-    float max_z, RenderData* render_data) {
-  for (int i = 0; i < options_.landmark_connections_size(); i += 2) {
-    const auto& ld0 = landmarks.landmark(options_.landmark_connections(i));
-    const auto& ld1 = landmarks.landmark(options_.landmark_connections(i + 1));
-    const int gray_val1 =
-        255 - static_cast<int>(Remap(ld0.z(), min_z, max_z, 255));
-    const int gray_val2 =
-        255 - static_cast<int>(Remap(ld1.z(), min_z, max_z, 255));
-    AddConnectionToRenderData(ld0.x(), ld0.y(), ld1.x(), ld1.y(), options_,
-                              normalized, gray_val1, gray_val2, render_data);
-  }
-}
-
-void LandmarksToRenderDataCalculator::AddConnectionToRenderData(
-    float start_x, float start_y, float end_x, float end_y,
-    const LandmarksToRenderDataCalculatorOptions& options, bool normalized,
-    int gray_val1, int gray_val2, RenderData* render_data) {
-  auto* connection_annotation = render_data->add_render_annotations();
-  RenderAnnotation::GradientLine* line =
-      connection_annotation->mutable_gradient_line();
-  line->set_x_start(start_x);
-  line->set_y_start(start_y);
-  line->set_x_end(end_x);
-  line->set_y_end(end_y);
-  line->set_normalized(normalized);
-  line->mutable_color1()->set_r(gray_val1);
-  line->mutable_color1()->set_g(gray_val1);
-  line->mutable_color1()->set_b(gray_val1);
-  line->mutable_color2()->set_r(gray_val2);
-  line->mutable_color2()->set_g(gray_val2);
-  line->mutable_color2()->set_b(gray_val2);
-  connection_annotation->set_thickness(options.thickness());
-}
-
-template <class LandmarkListType>
-void LandmarksToRenderDataCalculator::AddConnections(
-    const LandmarkListType& landmarks, bool normalized,
-    RenderData* render_data) {
-  for (int i = 0; i < options_.landmark_connections_size(); i += 2) {
-    const auto& ld0 = landmarks.landmark(options_.landmark_connections(i));
-    const auto& ld1 = landmarks.landmark(options_.landmark_connections(i + 1));
-    AddConnectionToRenderData(ld0.x(), ld0.y(), ld1.x(), ld1.y(), options_,
-                              normalized, render_data);
-  }
-}
-
-void LandmarksToRenderDataCalculator::AddConnectionToRenderData(
-    float start_x, float start_y, float end_x, float end_y,
-    const LandmarksToRenderDataCalculatorOptions& options, bool normalized,
-    RenderData* render_data) {
-  auto* connection_annotation = render_data->add_render_annotations();
-  RenderAnnotation::Line* line = connection_annotation->mutable_line();
-  line->set_x_start(start_x);
-  line->set_y_start(start_y);
-  line->set_x_end(end_x);
-  line->set_y_end(end_y);
-  line->set_normalized(normalized);
-  SetColor(connection_annotation, options.connection_color());
-  connection_annotation->set_thickness(options.thickness());
-}
-
-RenderAnnotation* LandmarksToRenderDataCalculator::AddPointRenderData(
-    const LandmarksToRenderDataCalculatorOptions& options,
-    RenderData* render_data) {
-  auto* landmark_data_annotation = render_data->add_render_annotations();
-  landmark_data_annotation->set_scene_tag(kLandmarkLabel);
-  SetRenderAnnotationColorThickness(options, landmark_data_annotation);
-  return landmark_data_annotation;
-}
-
-void LandmarksToRenderDataCalculator::SetRenderAnnotationColorThickness(
-    const LandmarksToRenderDataCalculatorOptions& options,
-    RenderAnnotation* render_annotation) {
-  SetColor(render_annotation, options.landmark_color());
-  render_annotation->set_thickness(options.thickness());
-}
-
 }  // namespace mediapipe

mediapipe/calculators/util/local_file_contents_calculator.cc

@@ -16,34 +16,80 @@
 #include <string>
 
 #include "mediapipe/framework/calculator_framework.h"
-#include "mediapipe/framework/port/file_helpers.h"
 #include "mediapipe/framework/port/status.h"
+#include "mediapipe/util/resource_util.h"
 
 namespace mediapipe {
+
+namespace {
+
+constexpr char kFilePathTag[] = "FILE_PATH";
+constexpr char kContentsTag[] = "CONTENTS";
+
+}  // namespace
+
 // The calculator takes the path to the local file as an input side packet and
 // outputs the contents of that file.
 //
+// NOTE: file loading can be batched by providing multiple input/output side
+// packets.
+//
 // Example config:
 // node {
 //   calculator: "LocalFileContentsCalculator"
 //   input_side_packet: "FILE_PATH:file_path"
 //   output_side_packet: "CONTENTS:contents"
 // }
+//
+// node {
+//   calculator: "LocalFileContentsCalculator"
+//   input_side_packet: "FILE_PATH:0:file_path1"
+//   input_side_packet: "FILE_PATH:1:file_path2"
+//   ...
+//   output_side_packet: "CONTENTS:0:contents1"
+//   output_side_packet: "CONTENTS:1:contents2"
+//   ...
+// }
 class LocalFileContentsCalculator : public CalculatorBase {
  public:
   static ::mediapipe::Status GetContract(CalculatorContract* cc) {
-    cc->InputSidePackets().Tag("FILE_PATH").Set<std::string>();
-    cc->OutputSidePackets().Tag("CONTENTS").Set<std::string>();
+    RET_CHECK(cc->InputSidePackets().HasTag(kFilePathTag))
+        << "Missing PATH input side packet(s)";
+    RET_CHECK(cc->OutputSidePackets().HasTag(kContentsTag))
+        << "Missing CONTENTS output side packet(s)";
+
+    RET_CHECK_EQ(cc->InputSidePackets().NumEntries(kFilePathTag),
+                 cc->OutputSidePackets().NumEntries(kContentsTag))
+        << "Same number of input streams and output streams is required.";
+
+    for (CollectionItemId id = cc->InputSidePackets().BeginId(kFilePathTag);
+         id != cc->InputSidePackets().EndId(kFilePathTag); ++id) {
+      cc->InputSidePackets().Get(id).Set<std::string>();
+    }
+
+    for (CollectionItemId id = cc->OutputSidePackets().BeginId(kContentsTag);
+         id != cc->OutputSidePackets().EndId(kContentsTag); ++id) {
+      cc->OutputSidePackets().Get(id).Set<std::string>();
+    }
+
     return ::mediapipe::OkStatus();
   }
 
   ::mediapipe::Status Open(CalculatorContext* cc) override {
-    std::string contents;
-    MP_RETURN_IF_ERROR(mediapipe::file::GetContents(
-        cc->InputSidePackets().Tag("FILE_PATH").Get<std::string>(), &contents));
-    cc->OutputSidePackets()
-        .Tag("CONTENTS")
-        .Set(MakePacket<std::string>(std::move(contents)));
+    CollectionItemId input_id = cc->InputSidePackets().BeginId(kFilePathTag);
+    CollectionItemId output_id = cc->OutputSidePackets().BeginId(kContentsTag);
+    // Number of inputs and outpus is the same according to the contract.
+    for (; input_id != cc->InputSidePackets().EndId(kFilePathTag);
+         ++input_id, ++output_id) {
+      std::string file_path =
+          cc->InputSidePackets().Get(input_id).Get<std::string>();
+      ASSIGN_OR_RETURN(file_path, PathToResourceAsFile(file_path));
+
+      std::string contents;
+      MP_RETURN_IF_ERROR(GetResourceContents(file_path, &contents));
+      cc->OutputSidePackets().Get(output_id).Set(
+          MakePacket<std::string>(std::move(contents)));
+    }
     return ::mediapipe::OkStatus();
   }
 

mediapipe/calculators/util/local_file_pattern_contents_calculator.cc

@@ -0,0 +1,75 @@
+// Copyright 2019 The MediaPipe Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <memory>
+#include <string>
+
+#include "mediapipe/framework/calculator_framework.h"
+#include "mediapipe/framework/port/file_helpers.h"
+#include "mediapipe/framework/port/status.h"
+
+namespace mediapipe {
+// The calculator takes the path to local directory and desired file suffix to
+// mach as input side packets, and outputs the contents of those files that
+// match the pattern. Those matched files will be sent sequentially through the
+// output stream with incremental timestamp difference by 1.
+//
+// Example config:
+// node {
+//   calculator: "LocalFilePatternContentsCalculator"
+//   input_side_packet: "FILE_DIRECTORY:file_directory"
+//   input_side_packet: "FILE_SUFFIX:file_suffix"
+//   output_stream: "CONTENTS:contents"
+// }
+class LocalFilePatternContentsCalculator : public CalculatorBase {
+ public:
+  static ::mediapipe::Status GetContract(CalculatorContract* cc) {
+    cc->InputSidePackets().Tag("FILE_DIRECTORY").Set<std::string>();
+    cc->InputSidePackets().Tag("FILE_SUFFIX").Set<std::string>();
+    cc->Outputs().Tag("CONTENTS").Set<std::string>();
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Open(CalculatorContext* cc) override {
+    MP_RETURN_IF_ERROR(::mediapipe::file::MatchFileTypeInDirectory(
+        cc->InputSidePackets().Tag("FILE_DIRECTORY").Get<std::string>(),
+        cc->InputSidePackets().Tag("FILE_SUFFIX").Get<std::string>(),
+        &filenames_));
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Process(CalculatorContext* cc) override {
+    if (current_output_ < filenames_.size()) {
+      auto contents = absl::make_unique<std::string>();
+      LOG(INFO) << filenames_[current_output_];
+      MP_RETURN_IF_ERROR(mediapipe::file::GetContents(
+          filenames_[current_output_], contents.get()));
+      ++current_output_;
+      cc->Outputs()
+          .Tag("CONTENTS")
+          .Add(contents.release(), Timestamp(current_output_));
+    } else {
+      return tool::StatusStop();
+    }
+    return ::mediapipe::OkStatus();
+  }
+
+ private:
+  std::vector<std::string> filenames_;
+  int current_output_ = 0;
+};
+
+REGISTER_CALCULATOR(LocalFilePatternContentsCalculator);
+
+}  // namespace mediapipe

mediapipe/calculators/util/packet_latency_calculator_test.cc

@@ -276,6 +276,7 @@ TEST_F(PacketLatencyCalculatorTest, DoesNotOutputUntilReferencePacketReceived) {
       "delayed_packet_0", Adopt(new double()).At(Timestamp(2))));
 
   // Send a reference packet with timestamp 10 usec.
+  simulation_clock_->Sleep(absl::Microseconds(1));
   MP_ASSERT_OK(graph_.AddPacketToInputStream(
       "camera_frames", Adopt(new double()).At(Timestamp(10))));
   simulation_clock_->Sleep(absl::Microseconds(1));

mediapipe/calculators/util/rect_to_render_data_calculator.cc

@@ -45,15 +45,17 @@ RenderAnnotation::Rectangle* NewRect(
 void SetRect(bool normalized, double xmin, double ymin, double width,
              double height, double rotation,
              RenderAnnotation::Rectangle* rect) {
-  if (xmin + width < 0.0 || ymin + height < 0.0) return;
-  if (normalized) {
-    if (xmin > 1.0 || ymin > 1.0) return;
+  if (rotation == 0.0) {
+    if (xmin + width < 0.0 || ymin + height < 0.0) return;
+    if (normalized) {
+      if (xmin > 1.0 || ymin > 1.0) return;
+    }
   }
   rect->set_normalized(normalized);
-  rect->set_left(normalized ? std::max(xmin, 0.0) : xmin);
-  rect->set_top(normalized ? std::max(ymin, 0.0) : ymin);
-  rect->set_right(normalized ? std::min(xmin + width, 1.0) : xmin + width);
-  rect->set_bottom(normalized ? std::min(ymin + height, 1.0) : ymin + height);
+  rect->set_left(xmin);
+  rect->set_top(ymin);
+  rect->set_right(xmin + width);
+  rect->set_bottom(ymin + height);
   rect->set_rotation(rotation);
 }
 

mediapipe/calculators/util/timed_box_list_id_to_label_calculator.cc

@@ -0,0 +1,105 @@
+// Copyright 2019 The MediaPipe Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "mediapipe/calculators/util/timed_box_list_id_to_label_calculator.pb.h"
+#include "mediapipe/framework/calculator_framework.h"
+#include "mediapipe/framework/packet.h"
+#include "mediapipe/framework/port/status.h"
+#include "mediapipe/util/resource_util.h"
+#include "mediapipe/util/tracking/box_tracker.pb.h"
+
+#if defined(MEDIAPIPE_MOBILE)
+#include "mediapipe/util/android/file/base/file.h"
+#include "mediapipe/util/android/file/base/helpers.h"
+#else
+#include "mediapipe/framework/port/file_helpers.h"
+#endif
+
+namespace mediapipe {
+
+using mediapipe::TimedBoxProto;
+using mediapipe::TimedBoxProtoList;
+
+// Takes a label map (from label IDs to names), and populate the label field in
+// TimedBoxProto according to it's ID.
+//
+// Example usage:
+// node {
+//   calculator: "TimedBoxListIdToLabelCalculator"
+//   input_stream: "input_timed_box_list"
+//   output_stream: "output_timed_box_list"
+//   node_options: {
+//     [mediapipe.TimedBoxListIdToLabelCalculatorOptions] {
+//       label_map_path: "labelmap.txt"
+//     }
+//   }
+// }
+class TimedBoxListIdToLabelCalculator : public CalculatorBase {
+ public:
+  static ::mediapipe::Status GetContract(CalculatorContract* cc);
+
+  ::mediapipe::Status Open(CalculatorContext* cc) override;
+  ::mediapipe::Status Process(CalculatorContext* cc) override;
+
+ private:
+  std::unordered_map<int, std::string> label_map_;
+};
+REGISTER_CALCULATOR(TimedBoxListIdToLabelCalculator);
+
+::mediapipe::Status TimedBoxListIdToLabelCalculator::GetContract(
+    CalculatorContract* cc) {
+  cc->Inputs().Index(0).Set<TimedBoxProtoList>();
+  cc->Outputs().Index(0).Set<TimedBoxProtoList>();
+
+  return ::mediapipe::OkStatus();
+}
+
+::mediapipe::Status TimedBoxListIdToLabelCalculator::Open(
+    CalculatorContext* cc) {
+  cc->SetOffset(TimestampDiff(0));
+
+  const auto& options =
+      cc->Options<::mediapipe::TimedBoxListIdToLabelCalculatorOptions>();
+
+  std::string string_path;
+  ASSIGN_OR_RETURN(string_path, PathToResourceAsFile(options.label_map_path()));
+  std::string label_map_string;
+  MP_RETURN_IF_ERROR(file::GetContents(string_path, &label_map_string));
+
+  std::istringstream stream(label_map_string);
+  std::string line;
+  int i = 0;
+  while (std::getline(stream, line)) {
+    label_map_[i++] = line;
+  }
+  return ::mediapipe::OkStatus();
+}
+
+::mediapipe::Status TimedBoxListIdToLabelCalculator::Process(
+    CalculatorContext* cc) {
+  const auto& input_list = cc->Inputs().Index(0).Get<TimedBoxProtoList>();
+  auto output_list = absl::make_unique<TimedBoxProtoList>();
+  for (const auto& input_box : input_list.box()) {
+    TimedBoxProto* box_ptr = output_list->add_box();
+    *box_ptr = input_box;
+
+    if (label_map_.find(input_box.id()) != label_map_.end()) {
+      box_ptr->set_label(label_map_[input_box.id()]);
+    }
+  }
+  cc->Outputs().Index(0).Add(output_list.release(), cc->InputTimestamp());
+  return ::mediapipe::OkStatus();
+}
+
+}  // namespace mediapipe

mediapipe/calculators/util/timed_box_list_id_to_label_calculator.proto

@@ -0,0 +1,28 @@
+// Copyright 2019 The MediaPipe Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+syntax = "proto2";
+
+package mediapipe;
+
+import "mediapipe/framework/calculator.proto";
+
+message TimedBoxListIdToLabelCalculatorOptions {
+  extend mediapipe.CalculatorOptions {
+    optional TimedBoxListIdToLabelCalculatorOptions ext = 297701606;
+  }
+
+  // Path to a label map file for getting the actual name of detected classes.
+  optional string label_map_path = 1;
+}

mediapipe/calculators/util/timed_box_list_to_render_data_calculator.cc

@@ -12,6 +12,8 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+#include <algorithm>
+
 #include "absl/memory/memory.h"
 #include "absl/strings/str_cat.h"
 #include "absl/strings/str_join.h"
@@ -66,6 +68,27 @@ void AddTimedBoxProtoToRenderData(
     rect->set_bottom(box_proto.bottom());
     rect->set_rotation(box_proto.rotation());
   }
+
+  if (box_proto.has_label()) {
+    auto* label_annotation = render_data->add_render_annotations();
+    label_annotation->mutable_color()->set_r(options.box_color().r());
+    label_annotation->mutable_color()->set_g(options.box_color().g());
+    label_annotation->mutable_color()->set_b(options.box_color().b());
+    label_annotation->set_thickness(options.thickness());
+    RenderAnnotation::Text* text = label_annotation->mutable_text();
+    text->set_display_text(box_proto.label());
+    text->set_normalized(true);
+    constexpr float text_left_start = 0.2f;
+    text->set_left((1.0f - text_left_start) * box_proto.left() +
+                   text_left_start * box_proto.right());
+    constexpr float text_baseline = 0.6f;
+    text->set_baseline(text_baseline * box_proto.bottom() +
+                       (1.0f - text_baseline) * box_proto.top());
+    constexpr float text_height = 0.1f;
+    text->set_font_height(std::min(box_proto.bottom() - box_proto.top(),
+                                   box_proto.right() - box_proto.left()) *
+                          text_height);
+  }
 }
 
 }  // namespace

mediapipe/calculators/video/BUILD

@@ -65,6 +65,26 @@ proto_library(
     ],
 )
 
+proto_library(
+    name = "tracked_detection_manager_calculator_proto",
+    srcs = ["tracked_detection_manager_calculator.proto"],
+    visibility = ["//visibility:public"],
+    deps = [
+        "//mediapipe/framework:calculator_proto",
+        "//mediapipe/util/tracking:tracked_detection_manager_config_proto",
+    ],
+)
+
+proto_library(
+    name = "box_detector_calculator_proto",
+    srcs = ["box_detector_calculator.proto"],
+    visibility = ["//visibility:public"],
+    deps = [
+        "//mediapipe/framework:calculator_proto",
+        "//mediapipe/util/tracking:box_detector_proto",
+    ],
+)
+
 proto_library(
     name = "video_pre_stream_calculator_proto",
     srcs = ["video_pre_stream_calculator.proto"],
@@ -107,6 +127,28 @@ mediapipe_cc_proto_library(
     deps = [":box_tracker_calculator_proto"],
 )
 
+mediapipe_cc_proto_library(
+    name = "tracked_detection_manager_calculator_cc_proto",
+    srcs = ["tracked_detection_manager_calculator.proto"],
+    cc_deps = [
+        "//mediapipe/framework:calculator_cc_proto",
+        "//mediapipe/util/tracking:tracked_detection_manager_config_cc_proto",
+    ],
+    visibility = ["//visibility:public"],
+    deps = [":tracked_detection_manager_calculator_proto"],
+)
+
+mediapipe_cc_proto_library(
+    name = "box_detector_calculator_cc_proto",
+    srcs = ["box_detector_calculator.proto"],
+    cc_deps = [
+        "//mediapipe/framework:calculator_cc_proto",
+        "//mediapipe/util/tracking:box_detector_cc_proto",
+    ],
+    visibility = ["//visibility:public"],
+    deps = [":box_detector_calculator_proto"],
+)
+
 mediapipe_cc_proto_library(
     name = "video_pre_stream_calculator_cc_proto",
     srcs = ["video_pre_stream_calculator.proto"],
@@ -138,7 +180,7 @@ cc_library(
     srcs = ["flow_to_image_calculator.cc"],
     visibility = ["//visibility:public"],
     deps = [
-        "//mediapipe/calculators/video:flow_to_image_calculator_cc_proto",
+        ":flow_to_image_calculator_cc_proto",
         "//mediapipe/calculators/video/tool:flow_quantizer_model",
         "//mediapipe/framework:calculator_framework",
         "//mediapipe/framework/formats:image_format_cc_proto",
@@ -279,11 +321,54 @@ cc_library(
     alwayslink = 1,
 )
 
+cc_library(
+    name = "box_detector_calculator",
+    srcs = ["box_detector_calculator.cc"],
+    visibility = ["//visibility:public"],
+    deps = [
+        ":box_detector_calculator_cc_proto",
+        "@com_google_absl//absl/memory",
+        "@com_google_absl//absl/strings",
+        "//mediapipe/framework:calculator_framework",
+        "//mediapipe/framework/formats:image_frame",
+        "//mediapipe/framework/formats:image_frame_opencv",
+        "//mediapipe/framework/formats:video_stream_header",  # fixdeps: keep -- required for exobazel build.
+        "//mediapipe/framework/port:integral_types",
+        "//mediapipe/framework/port:logging",
+        "//mediapipe/framework/port:opencv_core",
+        "//mediapipe/framework/port:opencv_features2d",
+        "//mediapipe/framework/port:ret_check",
+        "//mediapipe/framework/port:status",
+        "//mediapipe/util:resource_util",
+        "//mediapipe/util/tracking",
+        "//mediapipe/util/tracking:box_detector",
+        "//mediapipe/util/tracking:box_tracker",
+        "//mediapipe/util/tracking:box_tracker_cc_proto",
+        "//mediapipe/util/tracking:flow_packager_cc_proto",
+        "//mediapipe/util/tracking:tracking_visualization_utilities",
+    ] + select({
+        "//mediapipe:android": [
+            "//mediapipe/util/android/file/base",
+        ],
+        "//mediapipe:ios": [
+            "//mediapipe/util/android/file/base",
+        ],
+        "//mediapipe:macos": [
+            "//mediapipe/framework/port:file_helpers",
+        ],
+        "//conditions:default": [
+            "//mediapipe/framework/port:file_helpers",
+        ],
+    }),
+    alwayslink = 1,
+)
+
 cc_library(
     name = "tracked_detection_manager_calculator",
     srcs = ["tracked_detection_manager_calculator.cc"],
     visibility = ["//visibility:public"],
     deps = [
+        ":tracked_detection_manager_calculator_cc_proto",
         "//mediapipe/framework:calculator_framework",
         "//mediapipe/framework/formats:detection_cc_proto",
         "//mediapipe/framework/formats:location_data_cc_proto",
@@ -368,6 +453,7 @@ cc_test(
 cc_test(
     name = "tvl1_optical_flow_calculator_test",
     srcs = ["tvl1_optical_flow_calculator_test.cc"],
+    linkstatic = 1,
     deps = [
         ":tvl1_optical_flow_calculator",
         "//mediapipe/framework:calculator_framework",
@@ -383,20 +469,18 @@ cc_test(
     ],
 )
 
-MEDIAPIPE_DEPS = [
-    "//mediapipe/calculators/video:box_tracker_calculator",
-    "//mediapipe/calculators/video:flow_packager_calculator",
-    "//mediapipe/calculators/video:motion_analysis_calculator",
-    "//mediapipe/framework/stream_handler:fixed_size_input_stream_handler",
-    "//mediapipe/framework/stream_handler:sync_set_input_stream_handler",
-]
-
 mediapipe_binary_graph(
     name = "parallel_tracker_binarypb",
     graph = "testdata/parallel_tracker_graph.pbtxt",
     output_name = "testdata/parallel_tracker.binarypb",
     visibility = ["//visibility:public"],
-    deps = MEDIAPIPE_DEPS,
+    deps = [
+        ":box_tracker_calculator",
+        ":flow_packager_calculator",
+        ":motion_analysis_calculator",
+        "//mediapipe/framework/stream_handler:fixed_size_input_stream_handler",
+        "//mediapipe/framework/stream_handler:sync_set_input_stream_handler",
+    ],
 )
 
 mediapipe_binary_graph(
@@ -404,7 +488,13 @@ mediapipe_binary_graph(
     graph = "testdata/tracker_graph.pbtxt",
     output_name = "testdata/tracker.binarypb",
     visibility = ["//visibility:public"],
-    deps = MEDIAPIPE_DEPS,
+    deps = [
+        ":box_tracker_calculator",
+        ":flow_packager_calculator",
+        ":motion_analysis_calculator",
+        "//mediapipe/framework/stream_handler:fixed_size_input_stream_handler",
+        "//mediapipe/framework/stream_handler:sync_set_input_stream_handler",
+    ],
 )
 
 cc_test(

mediapipe/calculators/video/box_detector_calculator.cc

@@ -0,0 +1,393 @@
+// Copyright 2019 The MediaPipe Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <stdio.h>
+
+#include <memory>
+#include <unordered_set>
+
+#include "absl/memory/memory.h"
+#include "absl/strings/numbers.h"
+#include "mediapipe/calculators/video/box_detector_calculator.pb.h"
+#include "mediapipe/framework/calculator_framework.h"
+#include "mediapipe/framework/formats/image_frame.h"
+#include "mediapipe/framework/formats/image_frame_opencv.h"
+#include "mediapipe/framework/formats/video_stream_header.h"
+#include "mediapipe/framework/port/integral_types.h"
+#include "mediapipe/framework/port/logging.h"
+#include "mediapipe/framework/port/opencv_core_inc.h"
+#include "mediapipe/framework/port/opencv_features2d_inc.h"
+#include "mediapipe/framework/port/ret_check.h"
+#include "mediapipe/framework/port/status.h"
+#include "mediapipe/util/resource_util.h"
+#include "mediapipe/util/tracking/box_detector.h"
+#include "mediapipe/util/tracking/box_tracker.h"
+#include "mediapipe/util/tracking/box_tracker.pb.h"
+#include "mediapipe/util/tracking/flow_packager.pb.h"
+#include "mediapipe/util/tracking/tracking.h"
+#include "mediapipe/util/tracking/tracking_visualization_utilities.h"
+
+#if defined(MEDIAPIPE_MOBILE)
+#include "mediapipe/util/android/file/base/file.h"
+#include "mediapipe/util/android/file/base/helpers.h"
+#else
+#include "mediapipe/framework/port/file_helpers.h"
+#endif
+
+namespace mediapipe {
+
+// A calculator to detect reappeared box positions from single frame.
+//
+// Input stream:
+//   TRACKING: Input tracking data (proto TrackingData) containing features and
+//             descriptors.
+//   VIDEO:    Optional input video stream tracked boxes are rendered over
+//             (Required if VIZ is specified).
+//   FEATURES: Input feature points (std::vector<cv::KeyPoint>) in the original
+//             pixel space.
+//   DESCRIPTORS: Input feature descriptors (std::vector<float>). Actual feature
+//             dimension needs to be specified in detector_options.
+//   IMAGE_SIZE: Input image dimension.
+//   TRACKED_BOXES : input box tracking result (proto TimedBoxProtoList) from
+//             BoxTrackerCalculator.
+//   ADD_INDEX: Optional std::string containing binary format proto of type
+//             BoxDetectorIndex. Used for adding target index to the detector
+//             search index during runtime.
+//   CANCEL_OBJECT_ID: Optional id of box to be removed. This is recommended
+//             to be used with SyncSetInputStreamHandler.
+//   REACQ_SWITCH: Optional bool for swithcing on and off reacquisition
+//             functionality. User should initialize a graph with box detector
+//             calculator and be able to switch it on and off in runtime.
+//
+// Output streams:
+//   VIZ:   Optional output video stream with rendered box positions
+//          (requires VIDEO to be present)
+//   BOXES: Optional output stream of type TimedBoxProtoList for each lost box.
+//
+// Imput side packets:
+//   INDEX_PROTO_STRING: Optional std::string containing binary format proto of
+//   type
+//                       BoxDetectorIndex. Used for initializing box_detector
+//                       with predefined template images.
+//   FRAME_ALIGNMENT:    Optional integer to indicate alignment_boundary for
+//                       outputing ImageFrame in "VIZ" stream.
+//                       Set to ImageFrame::kDefaultAlignmentBoundary for
+//                       offline pipeline to be compatible with FFmpeg.
+//                       Set to ImageFrame::kGlDefaultAlignmentBoundary for Apps
+//                       to be compatible with GL renderer.
+//   OUTPUT_INDEX_FILENAME: File path to the output index file.
+
+class BoxDetectorCalculator : public CalculatorBase {
+ public:
+  ~BoxDetectorCalculator() override = default;
+
+  static ::mediapipe::Status GetContract(CalculatorContract* cc);
+
+  ::mediapipe::Status Open(CalculatorContext* cc) override;
+  ::mediapipe::Status Process(CalculatorContext* cc) override;
+  ::mediapipe::Status Close(CalculatorContext* cc) override;
+
+ private:
+  BoxDetectorCalculatorOptions options_;
+  std::unique_ptr<BoxDetectorInterface> box_detector_;
+  bool detector_switch_ = true;
+  uint32 frame_alignment_ = ImageFrame::kDefaultAlignmentBoundary;
+  bool write_index_ = false;
+  int box_id_ = 0;
+};
+
+REGISTER_CALCULATOR(BoxDetectorCalculator);
+
+::mediapipe::Status BoxDetectorCalculator::GetContract(CalculatorContract* cc) {
+  if (cc->Inputs().HasTag("TRACKING")) {
+    cc->Inputs().Tag("TRACKING").Set<TrackingData>();
+  }
+
+  if (cc->Inputs().HasTag("TRACKED_BOXES")) {
+    cc->Inputs().Tag("TRACKED_BOXES").Set<TimedBoxProtoList>();
+  }
+
+  if (cc->Inputs().HasTag("VIDEO")) {
+    cc->Inputs().Tag("VIDEO").Set<ImageFrame>();
+  }
+
+  if (cc->Inputs().HasTag("FEATURES")) {
+    RET_CHECK(cc->Inputs().HasTag("DESCRIPTORS"))
+        << "FEATURES and DESCRIPTORS need to be specified together.";
+    cc->Inputs().Tag("FEATURES").Set<std::vector<cv::KeyPoint>>();
+  }
+
+  if (cc->Inputs().HasTag("DESCRIPTORS")) {
+    RET_CHECK(cc->Inputs().HasTag("FEATURES"))
+        << "FEATURES and DESCRIPTORS need to be specified together.";
+    cc->Inputs().Tag("DESCRIPTORS").Set<std::vector<float>>();
+  }
+
+  if (cc->Inputs().HasTag("IMAGE_SIZE")) {
+    cc->Inputs().Tag("IMAGE_SIZE").Set<std::pair<int, int>>();
+  }
+
+  if (cc->Inputs().HasTag("ADD_INDEX")) {
+    cc->Inputs().Tag("ADD_INDEX").Set<std::string>();
+  }
+
+  if (cc->Inputs().HasTag("CANCEL_OBJECT_ID")) {
+    cc->Inputs().Tag("CANCEL_OBJECT_ID").Set<int>();
+  }
+
+  if (cc->Inputs().HasTag("REACQ_SWITCH")) {
+    cc->Inputs().Tag("REACQ_SWITCH").Set<bool>();
+  }
+
+  if (cc->Outputs().HasTag("BOXES")) {
+    cc->Outputs().Tag("BOXES").Set<TimedBoxProtoList>();
+  }
+
+  if (cc->Outputs().HasTag("VIZ")) {
+    RET_CHECK(cc->Inputs().HasTag("VIDEO"))
+        << "Output stream VIZ requires VIDEO to be present.";
+    cc->Outputs().Tag("VIZ").Set<ImageFrame>();
+  }
+
+  if (cc->InputSidePackets().HasTag("INDEX_PROTO_STRING")) {
+    cc->InputSidePackets().Tag("INDEX_PROTO_STRING").Set<std::string>();
+  }
+
+  if (cc->InputSidePackets().HasTag("OUTPUT_INDEX_FILENAME")) {
+    cc->InputSidePackets().Tag("OUTPUT_INDEX_FILENAME").Set<std::string>();
+  }
+
+  if (cc->InputSidePackets().HasTag("FRAME_ALIGNMENT")) {
+    cc->InputSidePackets().Tag("FRAME_ALIGNMENT").Set<int>();
+  }
+
+  return ::mediapipe::OkStatus();
+}
+
+::mediapipe::Status BoxDetectorCalculator::Open(CalculatorContext* cc) {
+  options_ = cc->Options<BoxDetectorCalculatorOptions>();
+  box_detector_ = BoxDetectorInterface::Create(options_.detector_options());
+
+  if (cc->InputSidePackets().HasTag("INDEX_PROTO_STRING")) {
+    BoxDetectorIndex predefined_index;
+    if (!predefined_index.ParseFromString(cc->InputSidePackets()
+                                              .Tag("INDEX_PROTO_STRING")
+                                              .Get<std::string>())) {
+      LOG(FATAL) << "failed to parse BoxDetectorIndex from INDEX_PROTO_STRING";
+    }
+    box_detector_->AddBoxDetectorIndex(predefined_index);
+  }
+
+  for (const auto& filename : options_.index_proto_filename()) {
+    std::string string_path;
+    ASSIGN_OR_RETURN(string_path, PathToResourceAsFile(filename));
+    std::string index_string;
+    MP_RETURN_IF_ERROR(file::GetContents(string_path, &index_string));
+    BoxDetectorIndex predefined_index;
+    if (!predefined_index.ParseFromString(index_string)) {
+      LOG(FATAL)
+          << "failed to parse BoxDetectorIndex from index_proto_filename";
+    }
+    box_detector_->AddBoxDetectorIndex(predefined_index);
+  }
+
+  if (cc->InputSidePackets().HasTag("OUTPUT_INDEX_FILENAME")) {
+    write_index_ = true;
+  }
+
+  if (cc->InputSidePackets().HasTag("FRAME_ALIGNMENT")) {
+    frame_alignment_ = cc->InputSidePackets().Tag("FRAME_ALIGNMENT").Get<int>();
+  }
+
+  return ::mediapipe::OkStatus();
+}
+
+::mediapipe::Status BoxDetectorCalculator::Process(CalculatorContext* cc) {
+  const Timestamp timestamp = cc->InputTimestamp();
+  const int64 timestamp_msec = timestamp.Value() / 1000;
+
+  InputStream* cancel_object_id_stream =
+      cc->Inputs().HasTag("CANCEL_OBJECT_ID")
+          ? &(cc->Inputs().Tag("CANCEL_OBJECT_ID"))
+          : nullptr;
+  if (cancel_object_id_stream && !cancel_object_id_stream->IsEmpty()) {
+    const int cancel_object_id = cancel_object_id_stream->Get<int>();
+    box_detector_->CancelBoxDetection(cancel_object_id);
+  }
+
+  InputStream* add_index_stream = cc->Inputs().HasTag("ADD_INDEX")
+                                      ? &(cc->Inputs().Tag("ADD_INDEX"))
+                                      : nullptr;
+  if (add_index_stream && !add_index_stream->IsEmpty()) {
+    BoxDetectorIndex predefined_index;
+    if (!predefined_index.ParseFromString(
+            add_index_stream->Get<std::string>())) {
+      LOG(FATAL) << "failed to parse BoxDetectorIndex from ADD_INDEX";
+    }
+    box_detector_->AddBoxDetectorIndex(predefined_index);
+  }
+
+  InputStream* reacq_switch_stream = cc->Inputs().HasTag("REACQ_SWITCH")
+                                         ? &(cc->Inputs().Tag("REACQ_SWITCH"))
+                                         : nullptr;
+  if (reacq_switch_stream && !reacq_switch_stream->IsEmpty()) {
+    detector_switch_ = reacq_switch_stream->Get<bool>();
+  }
+
+  if (!detector_switch_) {
+    return ::mediapipe::OkStatus();
+  }
+
+  InputStream* track_stream = cc->Inputs().HasTag("TRACKING")
+                                  ? &(cc->Inputs().Tag("TRACKING"))
+                                  : nullptr;
+  InputStream* video_stream =
+      cc->Inputs().HasTag("VIDEO") ? &(cc->Inputs().Tag("VIDEO")) : nullptr;
+  InputStream* feature_stream = cc->Inputs().HasTag("FEATURES")
+                                    ? &(cc->Inputs().Tag("FEATURES"))
+                                    : nullptr;
+  InputStream* descriptor_stream = cc->Inputs().HasTag("DESCRIPTORS")
+                                       ? &(cc->Inputs().Tag("DESCRIPTORS"))
+                                       : nullptr;
+
+  CHECK(track_stream != nullptr || video_stream != nullptr ||
+        (feature_stream != nullptr && descriptor_stream != nullptr))
+      << "One and only one of {tracking_data, input image frame, "
+         "feature/descriptor} need to be valid.";
+
+  InputStream* tracked_boxes_stream = cc->Inputs().HasTag("TRACKED_BOXES")
+                                          ? &(cc->Inputs().Tag("TRACKED_BOXES"))
+                                          : nullptr;
+  std::unique_ptr<TimedBoxProtoList> detected_boxes(new TimedBoxProtoList());
+
+  if (track_stream != nullptr) {
+    // Detect from tracking data
+    if (track_stream->IsEmpty()) {
+      return ::mediapipe::OkStatus();
+    }
+
+    const TrackingData& tracking_data = track_stream->Get<TrackingData>();
+
+    CHECK(tracked_boxes_stream != nullptr) << "tracked_boxes needed.";
+
+    const TimedBoxProtoList tracked_boxes =
+        tracked_boxes_stream->Get<TimedBoxProtoList>();
+
+    box_detector_->DetectAndAddBox(tracking_data, tracked_boxes, timestamp_msec,
+                                   detected_boxes.get());
+  } else if (video_stream != nullptr) {
+    // Detect from input frame
+    if (video_stream->IsEmpty()) {
+      return ::mediapipe::OkStatus();
+    }
+
+    TimedBoxProtoList tracked_boxes;
+    if (tracked_boxes_stream != nullptr && !tracked_boxes_stream->IsEmpty()) {
+      tracked_boxes = tracked_boxes_stream->Get<TimedBoxProtoList>();
+    }
+
+    // Just directly pass along the image frame data as-is for detection; we
+    // don't need to worry about conforming to a specific alignment here.
+    const cv::Mat input_view =
+        formats::MatView(&video_stream->Get<ImageFrame>());
+    box_detector_->DetectAndAddBox(input_view, tracked_boxes, timestamp_msec,
+                                   detected_boxes.get());
+  } else {
+    if (feature_stream->IsEmpty() || descriptor_stream->IsEmpty()) {
+      return ::mediapipe::OkStatus();
+    }
+
+    const auto& image_size =
+        cc->Inputs().Tag("IMAGE_SIZE").Get<std::pair<int, int>>();
+    float inv_scale = 1.0f / std::max(image_size.first, image_size.second);
+
+    TimedBoxProtoList tracked_boxes;
+    if (tracked_boxes_stream != nullptr && !tracked_boxes_stream->IsEmpty()) {
+      tracked_boxes = tracked_boxes_stream->Get<TimedBoxProtoList>();
+    } else if (write_index_) {
+      auto* box_ptr = tracked_boxes.add_box();
+      box_ptr->set_id(box_id_);
+      box_ptr->set_reacquisition(true);
+      box_ptr->set_aspect_ratio((float)image_size.first /
+                                (float)image_size.second);
+
+      box_ptr->mutable_quad()->add_vertices(0);
+      box_ptr->mutable_quad()->add_vertices(0);
+
+      box_ptr->mutable_quad()->add_vertices(0);
+      box_ptr->mutable_quad()->add_vertices(1);
+
+      box_ptr->mutable_quad()->add_vertices(1);
+      box_ptr->mutable_quad()->add_vertices(1);
+
+      box_ptr->mutable_quad()->add_vertices(1);
+      box_ptr->mutable_quad()->add_vertices(0);
+
+      ++box_id_;
+    }
+
+    const auto& features = feature_stream->Get<std::vector<cv::KeyPoint>>();
+    const int feature_size = features.size();
+    std::vector<Vector2_f> features_vec(feature_size);
+
+    const auto& descriptors = descriptor_stream->Get<std::vector<float>>();
+    const int dims = options_.detector_options().descriptor_dims();
+    CHECK_GE(descriptors.size(), feature_size * dims);
+    cv::Mat descriptors_mat(feature_size, dims, CV_32F);
+    for (int j = 0; j < feature_size; ++j) {
+      features_vec[j].Set(features[j].pt.x * inv_scale,
+                          features[j].pt.y * inv_scale);
+      for (int i = 0; i < dims; ++i) {
+        descriptors_mat.at<float>(j, i) = descriptors[j * dims + i];
+      }
+    }
+
+    box_detector_->DetectAndAddBoxFromFeatures(
+        features_vec, descriptors_mat, tracked_boxes, timestamp_msec,
+        image_size.first * inv_scale, image_size.second * inv_scale,
+        detected_boxes.get());
+  }
+
+  if (cc->Outputs().HasTag("VIZ")) {
+    cv::Mat viz_view;
+    std::unique_ptr<ImageFrame> viz_frame;
+    if (video_stream != nullptr && !video_stream->IsEmpty()) {
+      viz_frame = absl::make_unique<ImageFrame>();
+      viz_frame->CopyFrom(video_stream->Get<ImageFrame>(), frame_alignment_);
+      viz_view = formats::MatView(viz_frame.get());
+    }
+    for (const auto& box : detected_boxes->box()) {
+      RenderBox(box, &viz_view);
+    }
+    cc->Outputs().Tag("VIZ").Add(viz_frame.release(), timestamp);
+  }
+
+  if (cc->Outputs().HasTag("BOXES")) {
+    cc->Outputs().Tag("BOXES").Add(detected_boxes.release(), timestamp);
+  }
+
+  return ::mediapipe::OkStatus();
+}
+
+::mediapipe::Status BoxDetectorCalculator::Close(CalculatorContext* cc) {
+  if (write_index_) {
+    BoxDetectorIndex index = box_detector_->ObtainBoxDetectorIndex();
+    MEDIAPIPE_CHECK_OK(mediapipe::file::SetContents(
+        cc->InputSidePackets().Tag("OUTPUT_INDEX_FILENAME").Get<std::string>(),
+        index.SerializeAsString()));
+  }
+  return ::mediapipe::OkStatus();
+}
+
+}  // namespace mediapipe

mediapipe/calculators/video/box_detector_calculator.proto

@@ -0,0 +1,31 @@
+// Copyright 2019 The MediaPipe Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+syntax = "proto2";
+
+package mediapipe;
+
+import "mediapipe/framework/calculator.proto";
+import "mediapipe/util/tracking/box_detector.proto";
+
+message BoxDetectorCalculatorOptions {
+  extend CalculatorOptions {
+    optional BoxDetectorCalculatorOptions ext = 289746530;
+  }
+
+  optional BoxDetectorOptions detector_options = 1;
+
+  // File path to the template index files.
+  repeated string index_proto_filename = 2;
+}

mediapipe/calculators/video/tracked_detection_manager_calculator.cc

@@ -18,6 +18,7 @@
 #include <vector>
 
 #include "absl/container/node_hash_map.h"
+#include "mediapipe/calculators/video/tracked_detection_manager_calculator.pb.h"
 #include "mediapipe/framework/calculator_framework.h"
 #include "mediapipe/framework/formats/detection.pb.h"
 #include "mediapipe/framework/formats/location_data.pb.h"
@@ -139,6 +140,7 @@ Detection GetAxisAlignedDetectionFromTrackedDetection(
 class TrackedDetectionManagerCalculator : public CalculatorBase {
  public:
   static ::mediapipe::Status GetContract(CalculatorContract* cc);
+  ::mediapipe::Status Open(CalculatorContext* cc) override;
 
   ::mediapipe::Status Process(CalculatorContext* cc) override;
 
@@ -184,6 +186,15 @@ REGISTER_CALCULATOR(TrackedDetectionManagerCalculator);
   return ::mediapipe::OkStatus();
 }
 
+::mediapipe::Status TrackedDetectionManagerCalculator::Open(
+    CalculatorContext* cc) {
+  mediapipe::TrackedDetectionManagerCalculatorOptions options =
+      cc->Options<mediapipe::TrackedDetectionManagerCalculatorOptions>();
+  tracked_detection_manager_.SetConfig(
+      options.tracked_detection_manager_options());
+  return ::mediapipe::OkStatus();
+}
+
 ::mediapipe::Status TrackedDetectionManagerCalculator::Process(
     CalculatorContext* cc) {
   if (cc->Inputs().HasTag("TRACKING_BOXES")) {

mediapipe/calculators/video/tracked_detection_manager_calculator.proto

@@ -0,0 +1,28 @@
+// Copyright 2020 The MediaPipe Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+syntax = "proto2";
+
+package mediapipe;
+
+import "mediapipe/framework/calculator.proto";
+import "mediapipe/util/tracking/tracked_detection_manager_config.proto";
+
+message TrackedDetectionManagerCalculatorOptions {
+  extend CalculatorOptions {
+    optional TrackedDetectionManagerCalculatorOptions ext = 301970230;
+  }
+
+  optional TrackedDetectionManagerConfig tracked_detection_manager_options = 1;
+}

mediapipe/docs/autoflip.md

@@ -15,6 +15,9 @@ For overall context on AutoFlip, please read this
 
 Run the following command to build the AutoFlip pipeline:
 
+Note: AutoFlip currently only works with OpenCV 3 . Please verify your OpenCV
+version beforehand.
+
 ```bash
 bazel build -c opt --define MEDIAPIPE_DISABLE_GPU=1 mediapipe/examples/desktop/autoflip:run_autoflip
 ```

mediapipe/docs/examples.md

@@ -44,6 +44,14 @@ graphs can be easily adapted to run on CPU v.s. GPU.
 [Object Detection and Tracking with GPU](./object_tracking_mobile_gpu.md) illustrates how to
 use MediaPipe for object detection and tracking.
 
+### Objectron: 3D Object Detection and Tracking with GPU
+
+[MediaPipe Objectron is 3D Object Detection with GPU](./objectron_mobile_gpu.md)
+illustrates mobile real-time 3D object detection and tracking pipeline for every
+day objects like shoes and chairs
+
+*   [Android](./objectron_mobile_gpu.md)
+
 ### Face Detection with GPU
 
 [Face Detection with GPU](./face_detection_mobile_gpu.md) illustrates how to use

mediapipe/docs/face_detection_desktop.md

@@ -51,7 +51,7 @@ To build and run the TensorFlow Lite example on desktop (GPU) with Webcam, run:
 ```bash
 # Video from webcam running on desktop GPU
 # This works only for linux currently
-$ bazel build -c opt --copt -DMESA_EGL_NO_X11_HEADERS \
+$ bazel build -c opt --copt -DMESA_EGL_NO_X11_HEADERS --copt -DEGL_NO_X11 \
     mediapipe/examples/desktop/face_detection:face_detection_gpu
 
 # It should print:
@@ -259,9 +259,9 @@ node {
 # Draws annotations and overlays them on top of the input images.
 node {
   calculator: "AnnotationOverlayCalculator"
-  input_stream: "INPUT_FRAME:throttled_input_video"
+  input_stream: "IMAGE:throttled_input_video"
   input_stream: "render_data"
-  output_stream: "OUTPUT_FRAME:output_video"
+  output_stream: "IMAGE:output_video"
 }
 ```
 

mediapipe/docs/face_detection_mobile_cpu.md

@@ -229,9 +229,9 @@ node {
 # Draws annotations and overlays them on top of the input images.
 node {
   calculator: "AnnotationOverlayCalculator"
-  input_stream: "INPUT_FRAME:input_video_cpu"
+  input_stream: "IMAGE:input_video_cpu"
   input_stream: "render_data"
-  output_stream: "OUTPUT_FRAME:output_video_cpu"
+  output_stream: "IMAGE:output_video_cpu"
 }
 
 # Transfers the annotated image from CPU back to GPU memory, to be sent out of

mediapipe/docs/face_detection_mobile_gpu.md

@@ -221,8 +221,8 @@ node {
 # Draws annotations and overlays them on top of the input images.
 node {
   calculator: "AnnotationOverlayCalculator"
-  input_stream: "INPUT_FRAME_GPU:throttled_input_video"
+  input_stream: "IMAGE_GPU:throttled_input_video"
   input_stream: "render_data"
-  output_stream: "OUTPUT_FRAME_GPU:output_video"
+  output_stream: "IMAGE_GPU:output_video"
 }
 ```

mediapipe/docs/gpu.md

@@ -190,10 +190,10 @@ within the MediaPipe framework:
 bazel build --define MEDIAPIPE_DISABLE_GPU=1  <my-target>
 
 # to enable full GPU support (OpenGL ES 3.1+ & Metal)
-bazel build --copt -DMESA_EGL_NO_X11_HEADERS  <my-target>
+bazel build --copt -DMESA_EGL_NO_X11_HEADERS --copt -DEGL_NO_X11 <my-target>
 
 # to enable only OpenGL ES 3.0 and below (no GLES 3.1+ features)
-bazel build --copt -DMESA_EGL_NO_X11_HEADERS --copt -DMEDIAPIPE_DISABLE_GL_COMPUTE  <my-target>
+bazel build --copt -DMESA_EGL_NO_X11_HEADERS --copt -DEGL_NO_X11 --copt -DMEDIAPIPE_DISABLE_GL_COMPUTE  <my-target>
 ```
 
 Note *MEDIAPIPE_DISABLE_GL_COMPUTE* is automatically defined on all Apple

mediapipe/docs/hair_segmentation_desktop.md

@@ -26,7 +26,7 @@ To build and run the TensorFlow Lite example on desktop (GPU) with Webcam, run:
 ```bash
 # Video from webcam running on desktop GPU
 # This works only for linux currently
-$ bazel build -c opt --copt -DMESA_EGL_NO_X11_HEADERS \
+$ bazel build -c opt --copt -DMESA_EGL_NO_X11_HEADERS --copt -DEGL_NO_X11 \
     mediapipe/examples/desktop/hair_segmentation:hair_segmentation_gpu
 
 # It should print:

mediapipe/docs/hand_detection_mobile_gpu.md

@@ -136,10 +136,10 @@ node {
 # Draws annotations and overlays them on top of the input images.
 node {
   calculator: "AnnotationOverlayCalculator"
-  input_stream: "INPUT_FRAME_GPU:throttled_input_video"
+  input_stream: "IMAGE_GPU:throttled_input_video"
   input_stream: "detection_render_data"
   input_stream: "rect_render_data"
-  output_stream: "OUTPUT_FRAME_GPU:output_video"
+  output_stream: "IMAGE_GPU:output_video"
 }
 ```
 

mediapipe/docs/hand_tracking_desktop.md

@@ -48,7 +48,7 @@ To build and run the TensorFlow Lite example on desktop (GPU) with Webcam, run:
 ```bash
 # Video from webcam running on desktop GPU
 # This works only for linux currently
-$ bazel build -c opt --copt -DMESA_EGL_NO_X11_HEADERS \
+$ bazel build -c opt --copt -DMESA_EGL_NO_X11_HEADERS --copt -DEGL_NO_X11 \
     mediapipe/examples/desktop/hand_tracking:hand_tracking_gpu
 
 # It should print:

mediapipe/docs/hand_tracking_mobile_gpu.md

@@ -716,10 +716,10 @@ node {
 # Draws annotations and overlays them on top of the input images.
 node {
   calculator: "AnnotationOverlayCalculator"
-  input_stream: "INPUT_FRAME_GPU:input_image"
+  input_stream: "IMAGE_GPU:input_image"
   input_stream: "detection_render_data"
   input_stream: "landmark_render_data"
   input_stream: "rect_render_data"
-  output_stream: "OUTPUT_FRAME_GPU:output_image"
+  output_stream: "IMAGE_GPU:output_image"
 }
 ```

mediapipe/docs/install.md

@@ -40,12 +40,11 @@ To build and run iOS apps:
     $ cd mediapipe
     ```
 
-2.  Install Bazel (version between 0.24.1 and 1.2.1).
+2.  Install Bazel.
 
     Follow the official
     [Bazel documentation](https://docs.bazel.build/versions/master/install-ubuntu.html)
-    to install Bazel manually. Note that MediaPipe doesn't support Bazel 2.0.0+
-    yet.
+    to install Bazel 2.0 or higher.
 
 3.  Install OpenCV and FFmpeg.
 
@@ -111,7 +110,7 @@ To build and run iOS apps:
     # To compile with GPU support, replace
     --define MEDIAPIPE_DISABLE_GPU=1
     # with
-    --copt -DMESA_EGL_NO_X11_HEADERS
+    --copt -DMESA_EGL_NO_X11_HEADERS --copt -DEGL_NO_X11
     # when building GPU examples.
     ```
 
@@ -125,7 +124,7 @@ To build and run iOS apps:
         mediapipe/examples/desktop/hello_world:hello_world
 
     # If you are running on Linux desktop with GPU support enabled (via mesa drivers)
-    $ bazel run --copt -DMESA_EGL_NO_X11_HEADERS \
+    $ bazel run --copt -DMESA_EGL_NO_X11_HEADERS --copt -DEGL_NO_X11 \
         mediapipe/examples/desktop/hello_world:hello_world
 
     # Should print:
@@ -152,12 +151,11 @@ To build and run iOS apps:
     $ cd mediapipe
     ```
 
-2.  Install Bazel (version between 0.24.1 and 1.2.1).
+2.  Install Bazel.
 
     Follow the official
     [Bazel documentation](https://docs.bazel.build/versions/master/install-redhat.html)
-    to install Bazel manually. Note that MediaPipe doesn't support Bazel 2.0.0+
-    yet.
+    to install Bazel 2.0 or higher.
 
 3.  Install OpenCV.
 
@@ -241,23 +239,18 @@ To build and run iOS apps:
     $ cd mediapipe
     ```
 
-3.  Install Bazel (version between 0.24.1 and 1.1.0).
+3.  Install Bazel.
 
-    Option 1. Use package manager tool to install Bazel 1.1.0
+    Option 1. Use package manager tool to install Bazel
 
     ```bash
-    # If Bazel 1.1.0+ was installed.
-    $ brew uninstall bazel
-    # Install Bazel 1.1.0
-    $ brew install https://raw.githubusercontent.com/bazelbuild/homebrew-tap/f8a0fa981bcb1784a0d0823e14867b844e94fb3d/Formula/bazel.rb
-    $ brew link bazel
+    $ brew install bazel
     # Run 'bazel version' to check version of bazel
     ```
 
     Option 2. Follow the official
     [Bazel documentation](https://docs.bazel.build/versions/master/install-os-x.html#install-with-installer-mac-os-x)
-    to install any version of Bazel manually. Note that MediaPipe doesn't
-    support Bazel 1.1.0+ on macOS yet.
+    to install Bazel 2.0 or higher.
 
 4.  Install OpenCV and FFmpeg.
 
@@ -364,8 +357,10 @@ To build and run iOS apps:
 
 ### Installing on Windows Subsystem for Linux (WSL)
 
-Note: WSL has historically not provided access to USB cameras. Mediapipe can use
-a video file as input.
+Note: The pre-built OpenCV packages don't support cameras in WSL. Unless you
+[compile](https://funvision.blogspot.com/2019/12/opencv-web-camera-and-video-streams-in.html)
+OpenCV with FFMPEG and GStreamer in WSL, the live demos won't work with any
+cameras. Alternatively, you use a video file as input.
 
 1.  Follow the
     [instruction](https://docs.microsoft.com/en-us/windows/wsl/install-win10) to
@@ -373,7 +368,7 @@ a video file as input.
 
 2.  Install Windows ADB and start the ADB server in Windows.
 
-    Note: Window’s and WSL’s adb versions must be the same version, e.g., if WSL
+    Note: Windows' and WSL’s adb versions must be the same version, e.g., if WSL
     has ADB 1.0.39, you need to download the corresponding Windows ADB from
     [here](https://dl.google.com/android/repository/platform-tools_r26.0.1-windows.zip).
 
@@ -389,18 +384,18 @@ a video file as input.
     username@DESKTOP-TMVLBJ1:~$ sudo apt-get update && sudo apt-get install -y build-essential git python zip adb openjdk-8-jdk
     ```
 
-5.  Install Bazel (version between 0.24.1 and 1.2.1).
+5.  Install Bazel.
 
     ```bash
     username@DESKTOP-TMVLBJ1:~$ curl -sLO --retry 5 --retry-max-time 10 \
-    https://storage.googleapis.com/bazel/0.27.0/release/bazel-0.27.0-installer-linux-x86_64.sh && \
-    sudo mkdir -p /usr/local/bazel/0.27.0 && \
-    chmod 755 bazel-0.27.0-installer-linux-x86_64.sh && \
-    sudo ./bazel-0.27.0-installer-linux-x86_64.sh --prefix=/usr/local/bazel/0.27.0 && \
-    source /usr/local/bazel/0.27.0/lib/bazel/bin/bazel-complete.bash
+    https://storage.googleapis.com/bazel/2.0.0/release/bazel-2.0.0-installer-linux-x86_64.sh && \
+    sudo mkdir -p /usr/local/bazel/2.0.0 && \
+    chmod 755 bazel-2.0.0-installer-linux-x86_64.sh && \
+    sudo ./bazel-2.0.0-installer-linux-x86_64.sh --prefix=/usr/local/bazel/2.0.0 && \
+    source /usr/local/bazel/2.0.0/lib/bazel/bin/bazel-complete.bash
 
-    username@DESKTOP-TMVLBJ1:~$ /usr/local/bazel/0.27.0/lib/bazel/bin/bazel version && \
-    alias bazel='/usr/local/bazel/0.27.0/lib/bazel/bin/bazel'
+    username@DESKTOP-TMVLBJ1:~$ /usr/local/bazel/2.0.0/lib/bazel/bin/bazel version && \
+    alias bazel='/usr/local/bazel/2.0.0/lib/bazel/bin/bazel'
     ```
 
 6.  Checkout MediaPipe repository.

mediapipe/docs/media_sequence.md

@@ -89,11 +89,11 @@ process new data sets, in the documentation of
        dataset = d.as_dataset('test')
        # implement additional processing and batching here
        dataset_output = dataset.make_one_shot_iterator().get_next()
-       images = dataset_output=['images']
-       labels = dataset_output=['labels']
+       images = dataset_output['images']
+       labels = dataset_output['labels']
 
        with tf.Session() as sess:
-         images_, labels_ = sess.run(images, labels)
+         images_, labels_ = sess.run([images, labels])
        print('The shape of images_ is %s' % str(images_.shape))
        print('The shape of labels_ is %s' % str(labels_.shape))
     ```

mediapipe/docs/mediapipe_ios_setup.md

@@ -76,6 +76,10 @@ in the next section.
 
 2.  Open `mediapipe/Mediapipe.tulsiproj` using the Tulsi app.
 
+    Important: If Tulsi displays an error saying "Bazel could not be found",
+    press the "Bazel..." button in the Packages tab and select the `bazel`
+    executable in your homebrew `/bin/` directory.
+
 3.  Select the MediaPipe config in the Configs tab, then press the Generate
     button below. You will be asked for a location to save the Xcode project.
     Once the project is generated, it will be opened in Xcode.

mediapipe/docs/multi_hand_tracking_desktop.md

@@ -48,7 +48,7 @@ To build and run the TensorFlow Lite example on desktop (GPU) with Webcam, run:
 ```bash
 # Video from webcam running on desktop GPU
 # This works only for linux currently
-$ bazel build -c opt --copt -DMESA_EGL_NO_X11_HEADERS \
+$ bazel build -c opt --copt -DMESA_EGL_NO_X11_HEADERS --copt -DEGL_NO_X11 \
     mediapipe/examples/desktop/multi_hand_tracking:multi_hand_tracking_gpu
 
 # It should print:

mediapipe/docs/multi_hand_tracking_mobile_gpu.md

@@ -745,11 +745,11 @@ node {
 # a vector of RenderData objects and draws each of them on the input frame.
 node {
   calculator: "AnnotationOverlayCalculator"
-  input_stream: "INPUT_FRAME_GPU:input_image"
+  input_stream: "IMAGE_GPU:input_image"
   input_stream: "detection_render_data"
   input_stream: "multi_hand_rects_render_data"
   input_stream: "multi_palm_rects_render_data"
   input_stream: "VECTOR:0:multi_hand_landmarks_render_data"
-  output_stream: "OUTPUT_FRAME_GPU:output_image"
+  output_stream: "IMAGE_GPU:output_image"
 }
 ```

mediapipe/docs/object_detection_desktop.md

@@ -26,6 +26,7 @@ To build and run the TensorFlow example on desktop, run:
 $ bazel build -c opt \
     --define MEDIAPIPE_DISABLE_GPU=1 \
     --define no_aws_support=true \
+    --linkopt=-s \
     mediapipe/examples/desktop/object_detection:object_detection_tensorflow
 
 # It should print:
@@ -163,9 +164,9 @@ node {
 # the graph.
 node {
   calculator: "AnnotationOverlayCalculator"
-  input_stream: "INPUT_FRAME:input_video"
+  input_stream: "IMAGE:input_video"
   input_stream: "render_data"
-  output_stream: "OUTPUT_FRAME:output_video"
+  output_stream: "IMAGE:output_video"
 }
 
 # Encodes the annotated images into a video file, adopting properties specified
@@ -396,9 +397,9 @@ node {
 # the graph.
 node {
   calculator: "AnnotationOverlayCalculator"
-  input_stream: "INPUT_FRAME:input_video"
+  input_stream: "IMAGE:input_video"
   input_stream: "render_data"
-  output_stream: "OUTPUT_FRAME:output_video"
+  output_stream: "IMAGE:output_video"
 }
 
 # Encodes the annotated images into a video file, adopting properties specified

mediapipe/docs/object_detection_mobile_cpu.md

@@ -230,9 +230,9 @@ node {
 # Draws annotations and overlays them on top of the input images.
 node {
   calculator: "AnnotationOverlayCalculator"
-  input_stream: "INPUT_FRAME:throttled_input_video_cpu"
+  input_stream: "IMAGE:throttled_input_video_cpu"
   input_stream: "render_data"
-  output_stream: "OUTPUT_FRAME:output_video_cpu"
+  output_stream: "IMAGE:output_video_cpu"
 }
 
 # Transfers the annotated image from CPU back to GPU memory, to be sent out of

mediapipe/docs/object_detection_mobile_gpu.md

@@ -212,8 +212,8 @@ node {
 # Draws annotations and overlays them on top of the input images.
 node {
   calculator: "AnnotationOverlayCalculator"
-  input_stream: "INPUT_FRAME_GPU:throttled_input_video"
+  input_stream: "IMAGE_GPU:throttled_input_video"
   input_stream: "render_data"
-  output_stream: "OUTPUT_FRAME_GPU:output_video"
+  output_stream: "IMAGE_GPU:output_video"
 }
 ```

mediapipe/docs/object_tracking_mobile_gpu.md

@@ -467,9 +467,9 @@ node {
 # Draws annotations and overlays them on top of the input images.
 node {
   calculator: "AnnotationOverlayCalculator"
-  input_stream: "INPUT_FRAME_GPU:input_image"
+  input_stream: "IMAGE_GPU:input_image"
   input_stream: "detections_render_data"
-  output_stream: "OUTPUT_FRAME_GPU:output_image"
+  output_stream: "IMAGE_GPU:output_image"
 }
 
 ```
@@ -484,7 +484,8 @@ CPU.
 To build and run the app:
 
 ```bash
-bazel build -c opt mediapipe/examples/desktop/object_tracking:object_tracking_cpu
+bazel build -c opt mediapipe/examples/desktop/object_tracking:object_tracking_cpu \
+  --define MEDIAPIPE_DISABLE_GPU=1
 
 bazel-bin/mediapipe/examples/desktop/object_tracking/object_tracking_cpu \
   --calculator_graph_config_file=mediapipe/graphs/tracking/object_detection_tracking_desktop_live.pbtxt

mediapipe/docs/objectron_mobile_gpu.md

@@ -0,0 +1,489 @@
+# MediaPipe Objectron (GPU)
+
+This doc focuses on the
+[below example graph](https://github.com/google/mediapipe/tree/master/mediapipe/graphs/object_detection_3d/shoe_classic_occlusion_tracking.pbtxt)
+that performs 3D object detection and tracking with TensorFlow Lite on GPU.
+
+Objectron for shoes                                                             | Objectron for chairs
+:-----------------------------------------------------------------------------: | :------------------:
+![objectron_shoe_android_gpu_gif](images/mobile/objectron_shoe_android_gpu.gif) | ![objectron_chair_android_gpu_gif](images/mobile/objectron_chair_android_gpu.gif)
+
+For overall context on MediaPipe Objectron, please read the
+[Google AI Blog](https://mediapipe.page.link/objectron-aiblog). The Objectron's
+ML model (see also the [model card](https://mediapipe.page.link/objectron-mc))
+estimates a 3D bounding box for the detected object.
+
+## Android
+
+[Source](https://github.com/google/mediapipe/tree/master/mediapipe/examples/android/src/java/com/google/mediapipe/apps/objectdetection3d)
+
+An arm64 build of Objectron for shoes can be
+[downloaded here](https://drive.google.com/open?id=1S0K4hbWt3o31FfQ4QU3Rz7IHrvOUMx1d),
+and for chairs can be
+[downloaded here](https://drive.google.com/open?id=1MM8K-13bXLCVS1EHQ-KgkVyEahEPrKej).
+
+To build and install the Objectron for shoes:
+
+```bash
+bazel build -c opt --config android_arm64 mediapipe/examples/android/src/java/com/google/mediapipe/apps/objectdetection3d:objectdetection3d
+```
+
+Similarly to build and install the Objectron for chairs, add **--define
+chair=true** flag to build command.
+
+```bash
+bazel build -c opt --define chair=true --config android_arm64 mediapipe/examples/android/src/java/com/google/mediapipe/apps/objectdetection3d:objectdetection3d
+```
+
+Once the app is built, install in on Android device with:
+
+```bash
+adb install bazel-bin/mediapipe/examples/android/src/java/com/google/mediapipe/apps/objectdetection3d/objectdetection3d.apk
+```
+
+## Graph
+
+The Objectron main graph internally utilizes the Objectron detection subgraph,
+and the Objectron tracking subgraph. To visualize the graph as shown above, copy
+the text specification of the graph below and paste it into
+[MediaPipe Visualizer](https://viz.mediapipe.dev/).
+
+### Main Graph
+
+This is the main graph for the shoe detector. This graph runs detection and
+tracking and renders the output to the display.
+
+![object_detection_mobile_gpu_graph](images/mobile/object_detection_3d_android_gpu.png)
+
+[Source pbtxt file](https://github.com/google/mediapipe/tree/master/mediapipe/graphs/object_detection_3d/shoe_classic_occlusion_tracking.pbtxt)
+
+```bash
+# MediaPipe object detection 3D with tracking graph.
+
+# Images on GPU coming into and out of the graph.
+input_stream: "input_video"
+output_stream: "output_video"
+
+# Creates a copy of the input_video stream. At the end of the graph, the
+# GlAnimationOverlayCalculator will consume the input_video texture and draws
+# on top of it.
+node: {
+  calculator: "GlScalerCalculator"
+  input_stream: "VIDEO:input_video"
+  output_stream: "VIDEO:input_video_copy"
+}
+
+# Resamples the images by specific frame rate. This calculator is used to
+# control the frequecy of subsequent calculators/subgraphs, e.g. less power
+# consumption for expensive process.
+node {
+  calculator: "PacketResamplerCalculator"
+  input_stream: "DATA:input_video_copy"
+  output_stream: "DATA:sampled_input_video"
+  node_options: {
+    [type.googleapis.com/mediapipe.PacketResamplerCalculatorOptions] {
+      frame_rate: 5
+    }
+  }
+}
+
+node {
+  calculator: "ObjectronDetectionSubgraphGpu"
+  input_stream: "IMAGE_GPU:sampled_input_video"
+  output_stream: "ANNOTATIONS:objects"
+}
+
+node {
+  calculator: "ObjectronTrackingSubgraphGpu"
+  input_stream: "FRAME_ANNOTATION:objects"
+  input_stream: "IMAGE_GPU:input_video_copy"
+  output_stream: "LIFTED_FRAME_ANNOTATION:lifted_tracked_objects"
+}
+
+# The rendering nodes:
+# We are rendering two meshes: 1) a 3D bounding box, which we overlay directly
+# on the texture, and 2) a shoe CAD model, which we use as an occlusion mask.
+# These models are designed using different tools, so we supply a transformation
+# to bring both of them to the Objectron's coordinate system.
+
+# Creates a model matrices for the tracked object given the lifted 3D points.
+# This calculator does two things: 1) Estimates object's pose (orientation,
+# translation, and scale) from the 3D vertices, and
+# 2) bring the object from the objectron's coordinate system to the renderer
+# (OpenGL) coordinate system. Since the final goal is to render a mesh file on
+# top of the object, we also supply a transformation to bring the mesh to the
+# objectron's coordinate system, and rescale mesh to the unit size.
+node {
+  calculator: "AnnotationsToModelMatricesCalculator"
+  input_stream: "ANNOTATIONS:lifted_tracked_objects"
+  output_stream: "MODEL_MATRICES:model_matrices"
+  node_options: {
+    [type.googleapis.com/mediapipe.AnnotationsToModelMatricesCalculatorOptions] {
+      # Re-scale the CAD model to the size of a unit box
+      model_scale: 0.05
+      model_scale: 0.05
+      model_scale: 0.05
+      # Bring the box CAD model to objectron's coordinate system. This
+      # is equivalent of -pi/2 rotation along the y-axis (right-hand rule):
+      # Eigen::AngleAxisf(-M_PI / 2., Eigen::Vector3f::UnitY())
+      model_transformation: 0.0
+      model_transformation: 0.0
+      model_transformation: -1.0
+      model_transformation: 0.0
+      model_transformation: 0.0
+      model_transformation: 1.0
+      model_transformation: 0.0
+      model_transformation: 0.0
+      model_transformation: 1.0
+      model_transformation: 0.0
+      model_transformation: 0.0
+      model_transformation: 0.0
+      model_transformation: 0.0
+      model_transformation: 0.0
+      model_transformation: 0.0
+      model_transformation: 1.0
+    }
+  }
+}
+
+# Compute the model matrices for the CAD model of the shoe, to be used as an
+# occlusion mask. The model will be rendered at the exact same location as the
+# bounding box.
+node {
+  calculator: "AnnotationsToModelMatricesCalculator"
+  input_stream: "ANNOTATIONS:lifted_tracked_objects"
+  output_stream: "MODEL_MATRICES:mask_model_matrices"
+  #input_side_packet: "MODEL_SCALE:model_scale"
+  node_options: {
+    [type.googleapis.com/mediapipe.AnnotationsToModelMatricesCalculatorOptions] {
+      # Re-scale the CAD model to the size of a unit box
+      model_scale: 0.45
+      model_scale: 0.25
+      model_scale: 0.15
+      # Bring the shoe CAD model to Deep Pursuit 3D's coordinate system. This
+      # is equivalent of -pi/2 rotation along the x-axis (right-hand rule):
+      # Eigen::AngleAxisf(-M_PI / 2., Eigen::Vector3f::UnitX())
+      model_transformation: 1.0
+      model_transformation: 0.0
+      model_transformation: 0.0
+      model_transformation: 0.0
+      model_transformation: 0.0
+      model_transformation: 0.0
+      model_transformation: 1.0
+      model_transformation: 0.0
+      model_transformation: 0.0
+      model_transformation: -1.0
+      model_transformation: 0.0
+      model_transformation: 0.0
+      model_transformation: 0.0
+      model_transformation: 0.0
+      model_transformation: 0.0
+      model_transformation: 1.0
+    }
+  }
+}
+
+# Render everything together. First we render the 3D bounding box animation,
+# then we render the occlusion mask.
+node: {
+  calculator: "GlAnimationOverlayCalculator"
+  input_stream: "VIDEO:input_video"
+  input_stream: "MODEL_MATRICES:model_matrices"
+  input_stream: "MASK_MODEL_MATRICES:mask_model_matrices"
+  output_stream: "output_video"
+  input_side_packet: "TEXTURE:box_texture"
+  input_side_packet: "ANIMATION_ASSET:box_asset_name"
+  input_side_packet: "MASK_TEXTURE:obj_texture"
+  input_side_packet: "MASK_ASSET:obj_asset_name"
+  node_options: {
+    [type.googleapis.com/mediapipe.GlAnimationOverlayCalculatorOptions] {
+      # Output resolution is 480x640 with the aspect ratio of 0.75
+      aspect_ratio: 0.75
+      vertical_fov_degrees: 70.
+      animation_speed_fps: 25
+    }
+  }
+}
+
+```
+
+### Objectron Detection Subgraph
+
+Objectron detection subgraph uses the *TfLiteInferenceCalculator* to run
+inference and decodes the output tensor to *FrameAnnotation* protobuf. The
+*FrameAnnotation* contains nine keypoints: the bounding box's center, as well as
+its eight vertices. The boxes will be passed to the Objectron tracking subgraph.
+
+![object_detection_subgraph](images/mobile/objectron_detection_subgraph.png)
+
+[Source pbtxt file](https://github.com/google/mediapipe/tree/master/mediapipe/graphs/object_detection_3d/subgraphs/objectron_detection_gpu.pbtxt)
+
+```bash
+# MediaPipe Objectron detection gpu subgraph
+
+type: "ObjectronDetectionSubgraphGpu"
+
+input_stream: "IMAGE_GPU:input_video"
+output_stream: "ANNOTATIONS:objects"
+
+# Transforms the input image on GPU to a 480x640 image. To scale the input
+# image, the scale_mode option is set to FIT to preserve the aspect ratio,
+# resulting in potential letterboxing in the transformed image.
+node: {
+  calculator: "ImageTransformationCalculator"
+  input_stream: "IMAGE_GPU:input_video"
+  output_stream: "IMAGE_GPU:transformed_input_video"
+  node_options: {
+    [type.googleapis.com/mediapipe.ImageTransformationCalculatorOptions] {
+      output_width: 480
+      output_height: 640
+      scale_mode: FIT
+    }
+  }
+}
+
+# Converts the transformed input image on GPU into an image tensor stored as a
+# TfLiteTensor.
+node {
+  calculator: "TfLiteConverterCalculator"
+  input_stream: "IMAGE_GPU:transformed_input_video"
+  output_stream: "TENSORS_GPU:image_tensor"
+}
+
+# Generates a single side packet containing a TensorFlow Lite op resolver that
+# supports custom ops needed by the model used in this graph.
+node {
+  calculator: "TfLiteCustomOpResolverCalculator"
+  output_side_packet: "opresolver"
+  node_options: {
+    [type.googleapis.com/mediapipe.TfLiteCustomOpResolverCalculatorOptions] {
+      use_gpu: true
+    }
+  }
+}
+
+# Runs a TensorFlow Lite model on GPU that takes an image tensor and outputs a
+# vector of tensors representing, for instance, detection boxes/keypoints and
+# scores.
+node {
+  calculator: "TfLiteInferenceCalculator"
+  input_stream: "TENSORS_GPU:image_tensor"
+  output_stream: "TENSORS:detection_tensors"
+  input_side_packet: "CUSTOM_OP_RESOLVER:opresolver"
+  node_options: {
+    [type.googleapis.com/mediapipe.TfLiteInferenceCalculatorOptions] {
+      model_path: "object_detection_3d.tflite"
+    }
+  }
+}
+
+# Decodes the model's output tensor (the heatmap and the distance fields) to 2D
+# keypoints. There are nine 2D keypoints: one center keypoint and eight vertices
+# for the 3D bounding box. The calculator parameters determine's the decoder's
+# sensitivity.
+node {
+  calculator: "TfLiteTensorsToObjectsCalculator"
+  input_stream: "TENSORS:detection_tensors"
+  output_stream: "ANNOTATIONS:objects"
+  node_options: {
+    [type.googleapis.com/mediapipe.TfLiteTensorsToObjectsCalculatorOptions] {
+      num_classes: 1
+      num_keypoints: 9
+      decoder_config {
+        heatmap_threshold: 0.6
+        local_max_distance: 2
+        offset_scale_coef: 1.0
+        voting_radius: 2
+        voting_allowance: 1
+        voting_threshold: 0.2
+      }
+    }
+  }
+}
+```
+
+### Object Tracking Subgraph
+
+Object tracking subgraph uses a *BoxTracker* calculator which is a generic
+tracking library, also used in
+[Mediapipe's 2D Object Detection and Tracking](https://github.com/google/mediapipe/tree/master/mediapipe/g3doc/object_tracking_mobile_gpu.md).
+The tracking runs every frame and when a new detection is available, it
+consolidates the detection and tracking results. The tracker tracks the box with
+its 2D keypoints, so at the end we lift the 2D keypoints to 3D using EPnP
+algorithm in *Lift2DFrameAnnotationTo3D* Calculator.
+
+![object_tracking_subgraph](images/mobile/objectron_tracking_subgraph.png)
+
+[Source pbtxt file](https://github.com/google/mediapipe/tree/master/mediapipe/graphs/object_detection_3d/subgraphs/objectron_tracking_gpu.pbtxt)
+
+```bash
+# MediaPipe Objectron tracking gpu subgraph
+
+type: "ObjectronTrackingSubgraphGpu"
+
+input_stream: "FRAME_ANNOTATION:objects"
+input_stream: "IMAGE_GPU:input_video"
+output_stream: "LIFTED_FRAME_ANNOTATION:lifted_tracked_objects"
+
+
+# Converts the detected keypoints to Boxes, used by the tracking subgraph.
+node {
+  calculator: "FrameAnnotationToTimedBoxListCalculator"
+  input_stream: "FRAME_ANNOTATION:objects"
+  output_stream: "BOXES:start_pos"
+}
+
+node: {
+  calculator: "ImageTransformationCalculator"
+  input_stream: "IMAGE_GPU:input_video"
+  output_stream: "IMAGE_GPU:downscaled_input_video"
+  node_options: {
+    [type.googleapis.com/mediapipe.ImageTransformationCalculatorOptions] {
+      output_width: 240
+      output_height: 320
+    }
+  }
+}
+
+# Converts GPU buffer to ImageFrame for processing tracking.
+node: {
+  calculator: "GpuBufferToImageFrameCalculator"
+  input_stream: "downscaled_input_video"
+  output_stream: "downscaled_input_video_cpu"
+}
+
+# Performs motion analysis on an incoming video stream.
+node: {
+  calculator: "MotionAnalysisCalculator"
+  input_stream: "VIDEO:downscaled_input_video_cpu"
+  output_stream: "CAMERA:camera_motion"
+  output_stream: "FLOW:region_flow"
+
+  node_options: {
+    [type.googleapis.com/mediapipe.MotionAnalysisCalculatorOptions]: {
+      analysis_options {
+        analysis_policy: ANALYSIS_POLICY_CAMERA_MOBILE
+        flow_options {
+          fast_estimation_min_block_size: 100
+          top_inlier_sets: 1
+          frac_inlier_error_threshold: 3e-3
+          downsample_mode: DOWNSAMPLE_TO_INPUT_SIZE
+          verification_distance: 5.0
+          verify_long_feature_acceleration: true
+          verify_long_feature_trigger_ratio: 0.1
+          tracking_options {
+            max_features: 500
+            adaptive_extraction_levels: 2
+            min_eig_val_settings {
+              adaptive_lowest_quality_level: 2e-4
+            }
+            klt_tracker_implementation: KLT_OPENCV
+          }
+        }
+      }
+    }
+  }
+}
+
+# Reads optical flow fields defined in
+# mediapipe/framework/formats/motion/optical_flow_field.h,
+# returns a VideoFrame with 2 channels (v_x and v_y), each channel is quantized
+# to 0-255.
+node: {
+  calculator: "FlowPackagerCalculator"
+  input_stream: "FLOW:region_flow"
+  input_stream: "CAMERA:camera_motion"
+  output_stream: "TRACKING:tracking_data"
+
+  node_options: {
+    [type.googleapis.com/mediapipe.FlowPackagerCalculatorOptions]: {
+      flow_packager_options: {
+        binary_tracking_data_support: false
+      }
+    }
+  }
+}
+
+# Tracks box positions over time.
+node: {
+  calculator: "BoxTrackerCalculator"
+  input_stream: "TRACKING:tracking_data"
+  input_stream: "TRACK_TIME:input_video"
+  input_stream: "START_POS:start_pos"
+  input_stream: "CANCEL_OBJECT_ID:cancel_object_id"
+  input_stream_info: {
+    tag_index: "CANCEL_OBJECT_ID"
+    back_edge: true
+  }
+  output_stream: "BOXES:boxes"
+
+  input_stream_handler {
+    input_stream_handler: "SyncSetInputStreamHandler"
+    options {
+      [mediapipe.SyncSetInputStreamHandlerOptions.ext] {
+        sync_set {
+          tag_index: "TRACKING"
+          tag_index: "TRACK_TIME"
+        }
+        sync_set {
+          tag_index: "START_POS"
+        }
+        sync_set {
+          tag_index: "CANCEL_OBJECT_ID"
+        }
+      }
+    }
+  }
+
+  node_options: {
+    [type.googleapis.com/mediapipe.BoxTrackerCalculatorOptions]: {
+      tracker_options: {
+        track_step_options {
+          track_object_and_camera: true
+          tracking_degrees: TRACKING_DEGREE_OBJECT_ROTATION_SCALE
+          inlier_spring_force: 0.0
+          static_motion_temporal_ratio: 3e-2
+        }
+      }
+      visualize_tracking_data: false
+      streaming_track_data_cache_size: 100
+    }
+  }
+}
+
+# Consolidates tracking and detection results.
+node {
+  calculator: "FrameAnnotationTrackerCalculator"
+  input_stream: "FRAME_ANNOTATION:objects"
+  input_stream: "TRACKED_BOXES:boxes"
+  output_stream: "TRACKED_FRAME_ANNOTATION:tracked_objects"
+  output_stream: "CANCEL_OBJECT_ID:cancel_object_id"
+  node_options: {
+    [type.googleapis.com/mediapipe.FrameAnnotationTrackerCalculatorOptions] {
+      img_width: 240
+      img_height: 320
+    }
+  }
+
+  input_stream_handler {
+    input_stream_handler: "SyncSetInputStreamHandler"
+    options {
+      [mediapipe.SyncSetInputStreamHandlerOptions.ext] {
+        sync_set {
+          tag_index: "FRAME_ANNOTATION"
+        }
+        sync_set {
+          tag_index: "TRACKED_BOXES"
+        }
+      }
+    }
+  }
+}
+
+# Lift the tracked 2D keypoints to 3D using EPnP algorithm.
+node {
+  calculator: "Lift2DFrameAnnotationTo3DCalculator"
+  input_stream: "FRAME_ANNOTATION:tracked_objects"
+  output_stream: "LIFTED_FRAME_ANNOTATION:lifted_tracked_objects"
+}
+```

mediapipe/docs/youtube_8m.md

@@ -61,6 +61,8 @@ videos.
     ```bash
     # cd to the root directory of the MediaPipe repo
     cd -
+
+    pip3 install tf_slim
     python -m mediapipe.examples.desktop.youtube8m.generate_vggish_frozen_graph
     ```
 
@@ -78,7 +80,7 @@ videos.
 5.  Run the MediaPipe binary to extract the features.
 
     ```bash
-    bazel build -c opt \
+    bazel build -c opt --linkopt=-s \
       --define MEDIAPIPE_DISABLE_GPU=1 --define no_aws_support=true \
       mediapipe/examples/desktop/youtube8m:extract_yt8m_features
 
@@ -126,13 +128,13 @@ the inference for both local videos and the dataset
 2.  Build the inference binary.
 
     ```bash
-    bazel build -c opt --define='MEDIAPIPE_DISABLE_GPU=1' \
+    bazel build -c opt --define='MEDIAPIPE_DISABLE_GPU=1' --linkopt=-s \
       mediapipe/examples/desktop/youtube8m:model_inference
     ```
 
 3.  Run the python web server.
 
-    Note: pip install absl-py
+    Note: pip3 install absl-py
 
     ```bash
     python mediapipe/examples/desktop/youtube8m/viewer/server.py --root `pwd`
@@ -162,7 +164,7 @@ the inference for both local videos and the dataset
 3.  Build and run the inference binary.
 
     ```bash
-    bazel build -c opt --define='MEDIAPIPE_DISABLE_GPU=1' \
+    bazel build -c opt --define='MEDIAPIPE_DISABLE_GPU=1' --linkopt=-s \
       mediapipe/examples/desktop/youtube8m:model_inference
 
     # segment_size is the number of seconds window of frames.