Project import generated by Copybara.

GitOrigin-RevId: e3566e5029af25b0fc4b1071a49e49ae20aa5df6

.bazelrc

@@ -3,7 +3,7 @@
 # Basic build settings
 build --jobs 128
 build --define='absl=1'
-build --cxxopt='-std=c++11'
+build --cxxopt='-std=c++14'
 build --copt='-Wno-sign-compare'
 build --copt='-Wno-unused-function'
 build --copt='-Wno-uninitialized'

Dockerfile

@@ -30,6 +30,7 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
         unzip \
         python \
         python-pip \
+        python3-pip \
         libopencv-core-dev \
         libopencv-highgui-dev \
         libopencv-imgproc-dev \
@@ -42,9 +43,10 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
 
 RUN pip install --upgrade setuptools
 RUN pip install future
+RUN pip3 install six
 
 # Install bazel
-ARG BAZEL_VERSION=0.26.1
+ARG BAZEL_VERSION=1.1.0
 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

@@ -10,11 +10,13 @@
 ## ML Solutions in MediaPipe
 
 * [Hand Tracking](mediapipe/docs/hand_tracking_mobile_gpu.md)
+* [Multi-hand Tracking](mediapipe/docs/multi_hand_tracking_mobile_gpu.md)
 * [Face Detection](mediapipe/docs/face_detection_mobile_gpu.md)
 * [Hair Segmentation](mediapipe/docs/hair_segmentation_mobile_gpu.md)
 * [Object Detection](mediapipe/docs/object_detection_mobile_gpu.md)
 
 ![hand_tracking](mediapipe/docs/images/mobile/hand_tracking_3d_android_gpu_small.gif)
+![multi-hand_tracking](mediapipe/docs/images/mobile/multi_hand_tracking_android_gpu_small.gif)
 ![face_detection](mediapipe/docs/images/mobile/face_detection_android_gpu_small.gif)
 ![hair_segmentation](mediapipe/docs/images/mobile/hair_segmentation_android_gpu_small.gif)
 ![object_detection](mediapipe/docs/images/mobile/object_detection_android_gpu_small.gif)
@@ -23,7 +25,7 @@
 Follow these [instructions](mediapipe/docs/install.md).
 
 ## Getting started
-See mobile and desktop [examples](mediapipe/docs/examples.md).
+See mobile, desktop and Google Coral [examples](mediapipe/docs/examples.md).
 
 ## Documentation
 [MediaPipe Read-the-Docs](https://mediapipe.readthedocs.io/) or [docs.mediapipe.dev](https://docs.mediapipe.dev)
@@ -41,7 +43,11 @@ A web-based visualizer is hosted on [viz.mediapipe.dev](https://viz.mediapipe.de
 * [MediaPipe: A Framework for Building Perception Pipelines](https://arxiv.org/abs/1906.08172)
 
 ## Events
-* [ML Conference, Berlin 9-11 Dec 2019](https://mlconference.ai/machine-learning-advanced-development/mediapipe-building-real-time-cross-platform-mobile-web-edge-desktop-video-audio-ml-pipelines/)
+* [AI Nextcon 2020, 12-16 Feb 2020, Seattle](http://aisea20.xnextcon.com/)
+* [MediaPipe Madrid Meetup, 16 Dec 2019](https://www.meetup.com/Madrid-AI-Developers-Group/events/266329088/)
+* [MediaPipe London Meetup, Google 123 Building, 12 Dec 2019](https://www.meetup.com/London-AI-Tech-Talk/events/266329038)
+* [ML Conference, Berlin, 11 Dec 2019](https://mlconference.ai/machine-learning-advanced-development/mediapipe-building-real-time-cross-platform-mobile-web-edge-desktop-video-audio-ml-pipelines/)
+* [MediaPipe Berlin Meetup, Google Berlin, 11 Dec 2019](https://www.meetup.com/Berlin-AI-Tech-Talk/events/266328794/)
 * [The 3rd Workshop on YouTube-8M Large Scale Video Understanding Workshop](https://research.google.com/youtube8m/workshop2019/index.html) Seoul, Korea ICCV 2019
 * [AI DevWorld 2019](https://aidevworld.com) on Oct 10 in San Jose, California
 * [Google Industry Workshop at ICIP 2019](http://2019.ieeeicip.org/?action=page4&id=14#Google) [Presentation](https://docs.google.com/presentation/d/e/2PACX-1vRIBBbO_LO9v2YmvbHHEt1cwyqH6EjDxiILjuT0foXy1E7g6uyh4CesB2DkkEwlRDO9_lWfuKMZx98T/pub?start=false&loop=false&delayms=3000&slide=id.g556cc1a659_0_5) on Sept 24 in Taipei, Taiwan

WORKSPACE

@@ -10,20 +10,23 @@ http_archive(
     sha256 = "2ef429f5d7ce7111263289644d233707dba35e39696377ebab8b0bc701f7818e",
 )
 load("@bazel_skylib//lib:versions.bzl", "versions")
-versions.check(minimum_bazel_version = "0.24.1",
+versions.check(minimum_bazel_version = "0.24.1")
-               maximum_bazel_version = "0.29.1")
 
-# ABSL cpp library.
+# ABSL cpp library lts_2019_08_08.
 http_archive(
     name = "com_google_absl",
-    # Head commit on 2019-04-12.
-    # TODO: Switch to the latest absl version when the problem gets
-    # fixed.
     urls = [
-        "https://github.com/abseil/abseil-cpp/archive/a02f62f456f2c4a7ecf2be3104fe0c6e16fbad9a.tar.gz",
+        "https://github.com/abseil/abseil-cpp/archive/20190808.tar.gz",
     ],
-    sha256 = "d437920d1434c766d22e85773b899c77c672b8b4865d5dc2cd61a29fdff3cf03",
+    # Remove after https://github.com/abseil/abseil-cpp/issues/326 is solved.
-    strip_prefix = "abseil-cpp-a02f62f456f2c4a7ecf2be3104fe0c6e16fbad9a",
+    patches = [
+        "@//third_party:com_google_absl_f863b622fe13612433fdf43f76547d5edda0c93001.diff"
+    ],
+    patch_args = [
+        "-p1",
+    ],
+    strip_prefix = "abseil-cpp-20190808",
+    sha256 = "8100085dada279bf3ee00cd064d43b5f55e5d913be0dfe2906f06f8f28d5b37e"
 )
 
 http_archive(
@@ -104,9 +107,9 @@ http_archive(
     ],
 )
 
-# 2019-08-15
+# 2019-11-21
-_TENSORFLOW_GIT_COMMIT = "67def62936e28f97c16182dfcc467d8d1cae02b4"
+_TENSORFLOW_GIT_COMMIT = "f482488b481a799ca07e7e2d153cf47b8e91a60c"
-_TENSORFLOW_SHA256= "ddd4e3c056e7c0ff2ef29133b30fa62781dfbf8a903e99efb91a02d292fa9562"
+_TENSORFLOW_SHA256= "8d9118c2ce186c7e1403f04b96982fe72c184060c7f7a93e30a28dca358694f0"
 http_archive(
     name = "org_tensorflow",
     urls = [
@@ -115,13 +118,6 @@ http_archive(
     ],
     strip_prefix = "tensorflow-%s" % _TENSORFLOW_GIT_COMMIT,
     sha256 = _TENSORFLOW_SHA256,
-    patches = [
-        "@//third_party:tensorflow_065c20bf79253257c87bd4614bb9a7fdef015cbb.diff",
-        "@//third_party:tensorflow_f67fcbefce906cd419e4657f0d41e21019b71abd.diff",
-    ],
-    patch_args = [
-        "-p1",
-    ],
 )
 
 load("@org_tensorflow//tensorflow:workspace.bzl", "tf_workspace")
@@ -157,11 +153,10 @@ new_local_repository(
 
 http_archive(
     name = "android_opencv",
-    sha256 = "056b849842e4fa8751d09edbb64530cfa7a63c84ccd232d0ace330e27ba55d0b",
     build_file = "@//third_party:opencv_android.BUILD",
     strip_prefix = "OpenCV-android-sdk",
     type = "zip",
-    url = "https://github.com/opencv/opencv/releases/download/4.1.0/opencv-4.1.0-android-sdk.zip",
+    url = "https://github.com/opencv/opencv/releases/download/3.4.3/opencv-3.4.3-android-sdk.zip",
 )
 
 # After OpenCV 3.2.0, the pre-compiled opencv2.framework has google protobuf symbols, which will
@@ -192,13 +187,18 @@ 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"],
+    repositories = [
+        "https://dl.google.com/dl/android/maven2",
+        "https://repo1.maven.org/maven2",
+    ],
 )
 
 maven_server(
@@ -214,10 +214,10 @@ maven_jar(
 )
 
 maven_jar(
-     name = "androidx_concurrent_futures",
+    name = "androidx_concurrent_futures",
-     artifact = "androidx.concurrent:concurrent-futures:1.0.0-alpha03",
+    artifact = "androidx.concurrent:concurrent-futures:1.0.0-alpha03",
-     sha1 = "b528df95c7e2fefa2210c0c742bf3e491c1818ae",
+    sha1 = "b528df95c7e2fefa2210c0c742bf3e491c1818ae",
-     server = "google_server",
+    server = "google_server",
 )
 
 maven_jar(
@@ -255,18 +255,11 @@ android_sdk_repository(
 
 # iOS basic build deps.
 
-load("@bazel_tools//tools/build_defs/repo:git.bzl", "git_repository")
+http_archive(
-
-git_repository(
     name = "build_bazel_rules_apple",
-    remote = "https://github.com/bazelbuild/rules_apple.git",
+    sha256 = "bdc8e66e70b8a75da23b79f1f8c6207356df07d041d96d2189add7ee0780cf4e",
-    tag = "0.18.0",
+    strip_prefix = "rules_apple-b869b0d3868d78a1d4ffd866ccb304fb68aa12c3",
-    patches = [
+    url = "https://github.com/bazelbuild/rules_apple/archive/b869b0d3868d78a1d4ffd866ccb304fb68aa12c3.tar.gz",
-        "@//third_party:rules_apple_c0863d0596ae6b769a29fa3fb72ff036444fd249.diff",
-    ],
-    patch_args = [
-        "-p1",
-    ],
 )
 
 load(

mediapipe/calculators/audio/spectrogram_calculator.cc

@@ -113,8 +113,15 @@ class SpectrogramCalculator : public CalculatorBase {
   ::mediapipe::Status Close(CalculatorContext* cc) override;
 
  private:
-  Timestamp CurrentOutputTimestamp() {
+  Timestamp CurrentOutputTimestamp(CalculatorContext* cc) {
-    // Current output timestamp is the *center* of the next frame to be
+    if (use_local_timestamp_) {
+      return cc->InputTimestamp();
+    }
+    return CumulativeOutputTimestamp();
+  }
+
+  Timestamp CumulativeOutputTimestamp() {
+    // Cumulative output timestamp is the *center* of the next frame to be
     // emitted, hence delayed by half a window duration compared to relevant
     // input timestamp.
     return initial_input_timestamp_ +
@@ -141,6 +148,7 @@ class SpectrogramCalculator : public CalculatorBase {
       const OutputMatrixType postprocess_output_fn(const OutputMatrixType&),
       CalculatorContext* cc);
 
+  bool use_local_timestamp_;
   double input_sample_rate_;
   bool pad_final_packet_;
   int frame_duration_samples_;
@@ -173,6 +181,8 @@ const float SpectrogramCalculator::kLnPowerToDb = 4.342944819032518;
   SpectrogramCalculatorOptions spectrogram_options =
       cc->Options<SpectrogramCalculatorOptions>();
 
+  use_local_timestamp_ = spectrogram_options.use_local_timestamp();
+
   if (spectrogram_options.frame_duration_seconds() <= 0.0) {
     ::mediapipe::InvalidArgumentErrorBuilder(MEDIAPIPE_LOC)
         << "Invalid or missing frame_duration_seconds.\n"
@@ -351,11 +361,11 @@ template <class OutputMatrixType>
         << "Inconsistent number of spectrogram channels.";
     if (allow_multichannel_input_) {
       cc->Outputs().Index(0).Add(spectrogram_matrices.release(),
-                                 CurrentOutputTimestamp());
+                                 CurrentOutputTimestamp(cc));
     } else {
       cc->Outputs().Index(0).Add(
           new OutputMatrixType(spectrogram_matrices->at(0)),
-          CurrentOutputTimestamp());
+          CurrentOutputTimestamp(cc));
     }
     cumulative_completed_frames_ += output_vectors.size();
   }

mediapipe/calculators/audio/spectrogram_calculator.proto

@@ -66,4 +66,11 @@ message SpectrogramCalculatorOptions {
   // uniformly regardless of output type (i.e., even dBs are multiplied, not
   // offset).
   optional double output_scale = 7 [default = 1.0];
+
+  // If use_local_timestamp is true, the output packet's timestamp is based on
+  // the last sample of the packet and it's inferred from the latest input
+  // packet's timestamp.  If false, the output packet's timestamp is based on
+  // the cumulative timestamping, which is inferred from the intial input
+  // timestamp and the cumulative number of samples.
+  optional bool use_local_timestamp = 8 [default = false];
 }

mediapipe/calculators/core/BUILD

@@ -13,12 +13,12 @@
 # limitations under the License.
 #
 
+load("//mediapipe/framework/port:build_config.bzl", "mediapipe_cc_proto_library")
+
 licenses(["notice"])  # Apache 2.0
 
 package(default_visibility = ["//visibility:private"])
 
-load("//mediapipe/framework/port:build_config.bzl", "mediapipe_cc_proto_library")
-
 proto_library(
     name = "concatenate_vector_calculator_proto",
     srcs = ["concatenate_vector_calculator.proto"],
@@ -79,6 +79,13 @@ proto_library(
     ],
 )
 
+proto_library(
+    name = "clip_vector_size_calculator_proto",
+    srcs = ["clip_vector_size_calculator.proto"],
+    visibility = ["//visibility:public"],
+    deps = ["//mediapipe/framework:calculator_proto"],
+)
+
 mediapipe_cc_proto_library(
     name = "packet_cloner_calculator_cc_proto",
     srcs = ["packet_cloner_calculator.proto"],
@@ -111,6 +118,14 @@ mediapipe_cc_proto_library(
     deps = [":concatenate_vector_calculator_proto"],
 )
 
+mediapipe_cc_proto_library(
+    name = "clip_vector_size_calculator_cc_proto",
+    srcs = ["clip_vector_size_calculator.proto"],
+    cc_deps = ["//mediapipe/framework:calculator_cc_proto"],
+    visibility = ["//visibility:public"],
+    deps = [":clip_vector_size_calculator_proto"],
+)
+
 mediapipe_cc_proto_library(
     name = "dequantize_byte_array_calculator_cc_proto",
     srcs = ["dequantize_byte_array_calculator.proto"],
@@ -169,6 +184,66 @@ cc_test(
     ],
 )
 
+cc_library(
+    name = "begin_loop_calculator",
+    srcs = ["begin_loop_calculator.cc"],
+    hdrs = ["begin_loop_calculator.h"],
+    visibility = ["//visibility:public"],
+    deps = [
+        "//mediapipe/framework:calculator_context",
+        "//mediapipe/framework:calculator_contract",
+        "//mediapipe/framework:calculator_framework",
+        "//mediapipe/framework:collection_item_id",
+        "//mediapipe/framework:packet",
+        "//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",
+        "@com_google_absl//absl/memory",
+    ],
+    alwayslink = 1,
+)
+
+cc_library(
+    name = "end_loop_calculator",
+    srcs = ["end_loop_calculator.cc"],
+    hdrs = ["end_loop_calculator.h"],
+    visibility = ["//visibility:public"],
+    deps = [
+        "//mediapipe/framework:calculator_context",
+        "//mediapipe/framework:calculator_contract",
+        "//mediapipe/framework:calculator_framework",
+        "//mediapipe/framework:collection_item_id",
+        "//mediapipe/framework:packet",
+        "//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",
+    ],
+    alwayslink = 1,
+)
+
+cc_test(
+    name = "begin_end_loop_calculator_graph_test",
+    srcs = ["begin_end_loop_calculator_graph_test.cc"],
+    deps = [
+        ":begin_loop_calculator",
+        ":end_loop_calculator",
+        "//mediapipe/calculators/core:packet_cloner_calculator",
+        "//mediapipe/framework:calculator_context",
+        "//mediapipe/framework:calculator_contract",
+        "//mediapipe/framework:calculator_framework",
+        "//mediapipe/framework/port:gtest_main",
+        "//mediapipe/framework/port:integral_types",
+        "//mediapipe/framework/port:parse_text_proto",
+        "//mediapipe/framework/port:status",
+        "@com_google_absl//absl/memory",
+    ],
+)
+
 cc_library(
     name = "concatenate_vector_calculator",
     srcs = ["concatenate_vector_calculator.cc"],
@@ -219,6 +294,50 @@ cc_test(
     ],
 )
 
+cc_library(
+    name = "clip_vector_size_calculator",
+    srcs = ["clip_vector_size_calculator.cc"],
+    hdrs = ["clip_vector_size_calculator.h"],
+    visibility = ["//visibility:public"],
+    deps = [
+        ":clip_vector_size_calculator_cc_proto",
+        "//mediapipe/framework:calculator_framework",
+        "//mediapipe/framework/formats:rect_cc_proto",
+        "//mediapipe/framework/port:ret_check",
+        "//mediapipe/framework/port:status",
+        "@org_tensorflow//tensorflow/lite:framework",
+    ],
+    alwayslink = 1,
+)
+
+cc_library(
+    name = "clip_detection_vector_size_calculator",
+    srcs = ["clip_detection_vector_size_calculator.cc"],
+    visibility = ["//visibility:public"],
+    deps = [
+        ":clip_vector_size_calculator",
+        "//mediapipe/framework:calculator_framework",
+        "//mediapipe/framework/formats:detection_cc_proto",
+    ],
+    alwayslink = 1,
+)
+
+cc_test(
+    name = "clip_vector_size_calculator_test",
+    srcs = ["clip_vector_size_calculator_test.cc"],
+    deps = [
+        ":clip_vector_size_calculator",
+        "//mediapipe/calculators/core:packet_resampler_calculator_cc_proto",
+        "//mediapipe/framework:calculator_framework",
+        "//mediapipe/framework:calculator_runner",
+        "//mediapipe/framework:timestamp",
+        "//mediapipe/framework/port:gtest_main",
+        "//mediapipe/framework/port:parse_text_proto",
+        "//mediapipe/framework/port:status",
+        "@com_google_absl//absl/strings",
+    ],
+)
+
 cc_library(
     name = "counting_source_calculator",
     srcs = ["counting_source_calculator.cc"],
@@ -300,7 +419,7 @@ cc_library(
         "//visibility:public",
     ],
     deps = [
-        "//mediapipe/calculators/core:packet_cloner_calculator_cc_proto",
+        ":packet_cloner_calculator_cc_proto",
         "//mediapipe/framework:calculator_framework",
         "@com_google_absl//absl/strings",
     ],
@@ -572,6 +691,7 @@ cc_library(
         ":split_vector_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",
         "//mediapipe/util:resource_util",
@@ -761,3 +881,29 @@ cc_test(
         "//mediapipe/framework/port:status",
     ],
 )
+
+cc_library(
+    name = "stream_to_side_packet_calculator",
+    srcs = ["stream_to_side_packet_calculator.cc"],
+    visibility = ["//visibility:public"],
+    deps = [
+        "//mediapipe/framework:calculator_framework",
+        "//mediapipe/framework:timestamp",
+        "//mediapipe/framework/port:status",
+    ],
+    alwayslink = 1,
+)
+
+cc_test(
+    name = "stream_to_side_packet_calculator_test",
+    srcs = ["stream_to_side_packet_calculator_test.cc"],
+    deps = [
+        ":stream_to_side_packet_calculator",
+        "//mediapipe/framework:calculator_runner",
+        "//mediapipe/framework:packet",
+        "//mediapipe/framework:timestamp",
+        "//mediapipe/framework/port:gtest_main",
+        "//mediapipe/framework/port:status",
+        "@com_google_absl//absl/memory",
+    ],
+)

mediapipe/calculators/core/begin_end_loop_calculator_graph_test.cc

@@ -0,0 +1,335 @@
+// 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 <string>
+#include <vector>
+
+#include "absl/memory/memory.h"
+#include "mediapipe/calculators/core/begin_loop_calculator.h"
+#include "mediapipe/calculators/core/end_loop_calculator.h"
+#include "mediapipe/framework/calculator_contract.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_matchers.h"  // NOLINT
+
+namespace mediapipe {
+namespace {
+
+typedef BeginLoopCalculator<std::vector<int>> BeginLoopIntegerCalculator;
+REGISTER_CALCULATOR(BeginLoopIntegerCalculator);
+
+class IncrementCalculator : public CalculatorBase {
+ public:
+  static ::mediapipe::Status GetContract(CalculatorContract* cc) {
+    cc->Inputs().Index(0).Set<int>();
+    cc->Outputs().Index(0).Set<int>();
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Open(CalculatorContext* cc) override {
+    cc->SetOffset(TimestampDiff(0));
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Process(CalculatorContext* cc) override {
+    const int& input_int = cc->Inputs().Index(0).Get<int>();
+    auto output_int = absl::make_unique<int>(input_int + 1);
+    cc->Outputs().Index(0).Add(output_int.release(), cc->InputTimestamp());
+    return ::mediapipe::OkStatus();
+  }
+};
+
+REGISTER_CALCULATOR(IncrementCalculator);
+
+typedef EndLoopCalculator<std::vector<int>> EndLoopIntegersCalculator;
+REGISTER_CALCULATOR(EndLoopIntegersCalculator);
+
+class BeginEndLoopCalculatorGraphTest : public ::testing::Test {
+ protected:
+  BeginEndLoopCalculatorGraphTest() {
+    graph_config_ = ParseTextProtoOrDie<CalculatorGraphConfig>(
+        R"(
+          num_threads: 4
+          input_stream: "ints"
+          node {
+            calculator: "BeginLoopIntegerCalculator"
+            input_stream: "ITERABLE:ints"
+            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"
+          }
+        )");
+    tool::AddVectorSink("ints_plus_one", &graph_config_, &output_packets_);
+  }
+
+  CalculatorGraphConfig graph_config_;
+  std::vector<Packet> output_packets_;
+};
+
+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());
+
+  // EndLoopCalc will forward the timestamp bound because there are no elements
+  // in collection to output.
+  ASSERT_EQ(0, output_packets_.size());
+
+  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());
+
+  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>>());
+
+  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)));
+
+  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)));
+
+  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)));
+
+  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>>());
+
+  // 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 = {3, 4};
+  EXPECT_EQ(expected_output_vector2,
+            output_packets_[1].Get<std::vector<int>>());
+}
+
+class MultiplierCalculator : public CalculatorBase {
+ public:
+  static ::mediapipe::Status GetContract(CalculatorContract* cc) {
+    cc->Inputs().Index(0).Set<int>();
+    cc->Inputs().Index(1).Set<int>();
+    cc->Outputs().Index(0).Set<int>();
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Open(CalculatorContext* cc) override {
+    cc->SetOffset(TimestampDiff(0));
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Process(CalculatorContext* cc) override {
+    const int& input_int = cc->Inputs().Index(0).Get<int>();
+    const int& multiplier_int = cc->Inputs().Index(1).Get<int>();
+    auto output_int = absl::make_unique<int>(input_int * multiplier_int);
+    cc->Outputs().Index(0).Add(output_int.release(), cc->InputTimestamp());
+    return ::mediapipe::OkStatus();
+  }
+};
+
+REGISTER_CALCULATOR(MultiplierCalculator);
+
+class BeginEndLoopCalculatorGraphWithClonedInputsTest : public ::testing::Test {
+ protected:
+  BeginEndLoopCalculatorGraphWithClonedInputsTest() {
+    graph_config_ = ParseTextProtoOrDie<CalculatorGraphConfig>(
+        R"(
+          num_threads: 4
+          input_stream: "ints"
+          input_stream: "multiplier"
+          node {
+            calculator: "BeginLoopIntegerCalculator"
+            input_stream: "ITERABLE:ints"
+            input_stream: "CLONE:multiplier"
+            output_stream: "ITEM:int_at_loop"
+            output_stream: "CLONE:multiplier_cloned_at_loop"
+            output_stream: "BATCH_END:timestamp"
+          }
+          node {
+            calculator: "MultiplierCalculator"
+            input_stream: "int_at_loop"
+            input_stream: "multiplier_cloned_at_loop"
+            output_stream: "multiplied_int_at_loop"
+          }
+          node {
+            calculator: "EndLoopIntegersCalculator"
+            input_stream: "ITEM:multiplied_int_at_loop"
+            input_stream: "BATCH_END:timestamp"
+            output_stream: "ITERABLE:multiplied_ints"
+          }
+        )");
+    tool::AddVectorSink("multiplied_ints", &graph_config_, &output_packets_);
+  }
+
+  CalculatorGraphConfig graph_config_;
+  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>>();
+  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());
+
+  // EndLoopCalc will forward the timestamp bound because there are no elements
+  // in collection to output.
+  ASSERT_EQ(0, output_packets_.size());
+
+  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());
+
+  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>>());
+
+  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)));
+
+  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)));
+
+  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)));
+
+  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>>());
+
+  // 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>>());
+}
+
+}  // namespace
+}  // namespace mediapipe

mediapipe/calculators/core/begin_loop_calculator.cc

@@ -0,0 +1,34 @@
+// 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/core/begin_loop_calculator.h"
+
+#include <vector>
+
+#include "mediapipe/framework/formats/landmark.pb.h"
+#include "mediapipe/framework/formats/rect.pb.h"
+
+namespace mediapipe {
+
+// A calculator to process std::vector<NormalizedLandmarkList>.
+typedef BeginLoopCalculator<std::vector<::mediapipe::NormalizedLandmarkList>>
+    BeginLoopNormalizedLandmarkListVectorCalculator;
+REGISTER_CALCULATOR(BeginLoopNormalizedLandmarkListVectorCalculator);
+
+// A calculator to process std::vector<NormalizedRect>.
+typedef BeginLoopCalculator<std::vector<::mediapipe::NormalizedRect>>
+    BeginLoopNormalizedRectCalculator;
+REGISTER_CALCULATOR(BeginLoopNormalizedRectCalculator);
+
+}  // namespace mediapipe

mediapipe/calculators/core/begin_loop_calculator.h

@@ -0,0 +1,157 @@
+// 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.
+
+#ifndef MEDIAPIPE_CALCULATORS_CORE_BEGIN_LOOP_CALCULATOR_H_
+#define MEDIAPIPE_CALCULATORS_CORE_BEGIN_LOOP_CALCULATOR_H_
+
+#include "absl/memory/memory.h"
+#include "mediapipe/framework/calculator_context.h"
+#include "mediapipe/framework/calculator_contract.h"
+#include "mediapipe/framework/calculator_framework.h"
+#include "mediapipe/framework/collection_item_id.h"
+#include "mediapipe/framework/packet.h"
+#include "mediapipe/framework/port/integral_types.h"
+#include "mediapipe/framework/port/ret_check.h"
+#include "mediapipe/framework/port/status.h"
+
+namespace mediapipe {
+
+// Calculator for implementing loops on iterable collections inside a MediaPipe
+// graph.
+//
+// It is designed to be used like:
+//
+// node {
+//   calculator:    "BeginLoopWithIterableCalculator"
+//   input_stream:  "ITERABLE:input_iterable"      # IterableT @ext_ts
+//   output_stream: "ITEM:input_element"           # ItemT     @loop_internal_ts
+//   output_stream: "BATCH_END:ext_ts"             # Timestamp @loop_internal_ts
+// }
+//
+// node {
+//   calculator:    "ElementToBlaConverterSubgraph"
+//   input_stream:  "ITEM:input_to_loop_body"      # ItemT     @loop_internal_ts
+//   output_stream: "BLA:output_of_loop_body"      # ItemU     @loop_internal_ts
+// }
+//
+// node {
+//   calculator:    "EndLoopWithOutputCalculator"
+//   input_stream:  "ITEM:output_of_loop_body"     # ItemU     @loop_internal_ts
+//   input_stream:  "BATCH_END:ext_ts"             # Timestamp @loop_internal_ts
+//   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.
+template <typename IterableT>
+class BeginLoopCalculator : public CalculatorBase {
+  using ItemT = typename IterableT::value_type;
+
+ public:
+  static ::mediapipe::Status GetContract(CalculatorContract* cc) {
+    // A non-empty packet in the optional "TICK" input stream wakes up the
+    // calculator.
+    if (cc->Inputs().HasTag("TICK")) {
+      cc->Inputs().Tag("TICK").SetAny();
+    }
+
+    // An iterable collection in the input stream.
+    RET_CHECK(cc->Inputs().HasTag("ITERABLE"));
+    cc->Inputs().Tag("ITERABLE").Set<IterableT>();
+
+    // An element from the collection.
+    RET_CHECK(cc->Outputs().HasTag("ITEM"));
+    cc->Outputs().Tag("ITEM").Set<ItemT>();
+
+    RET_CHECK(cc->Outputs().HasTag("BATCH_END"));
+    cc->Outputs()
+        .Tag("BATCH_END")
+        .Set<Timestamp>(
+            // A flush signal to the corresponding EndLoopCalculator for it to
+            // emit the aggregated result with the timestamp contained in this
+            // flush signal packet.
+        );
+
+    // Input streams tagged with "CLONE" are cloned to the corresponding
+    // "CLONE" output streams at loop timestamps.
+    RET_CHECK(cc->Inputs().NumEntries("CLONE") ==
+              cc->Outputs().NumEntries("CLONE"));
+    if (cc->Inputs().NumEntries("CLONE") > 0) {
+      for (int i = 0; i < cc->Inputs().NumEntries("CLONE"); ++i) {
+        cc->Inputs().Get("CLONE", i).SetAny();
+        cc->Outputs().Get("CLONE", i).SetSameAs(&cc->Inputs().Get("CLONE", i));
+      }
+    }
+
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Process(CalculatorContext* cc) final {
+    Timestamp last_timestamp = loop_internal_timestamp_;
+    if (!cc->Inputs().Tag("ITERABLE").IsEmpty()) {
+      const IterableT& collection =
+          cc->Inputs().Tag("ITERABLE").template Get<IterableT>();
+      for (const auto& item : collection) {
+        cc->Outputs().Tag("ITEM").AddPacket(
+            MakePacket<ItemT>(item).At(loop_internal_timestamp_));
+        ForwardClonePackets(cc, loop_internal_timestamp_);
+        ++loop_internal_timestamp_;
+      }
+    }
+
+    // The collection was empty and nothing was processed.
+    if (last_timestamp == loop_internal_timestamp_) {
+      // Increment loop_internal_timestamp_ because it is used up now.
+      ++loop_internal_timestamp_;
+      for (auto it = cc->Outputs().begin(); it < cc->Outputs().end(); ++it) {
+        it->SetNextTimestampBound(loop_internal_timestamp_);
+      }
+    }
+
+    // The for loop processing the input collection already incremented
+    // loop_internal_timestamp_. To emit BATCH_END packet along the last
+    // non-BATCH_END packet, decrement by one.
+    cc->Outputs()
+        .Tag("BATCH_END")
+        .AddPacket(MakePacket<Timestamp>(cc->InputTimestamp())
+                       .At(Timestamp(loop_internal_timestamp_ - 1)));
+
+    return ::mediapipe::OkStatus();
+  }
+
+ private:
+  void ForwardClonePackets(CalculatorContext* cc, Timestamp output_timestamp) {
+    if (cc->Inputs().NumEntries("CLONE") > 0) {
+      for (int i = 0; i < cc->Inputs().NumEntries("CLONE"); ++i) {
+        if (!cc->Inputs().Get("CLONE", i).IsEmpty()) {
+          auto input_packet = cc->Inputs().Get("CLONE", i).Value();
+          cc->Outputs()
+              .Get("CLONE", i)
+              .AddPacket(std::move(input_packet).At(output_timestamp));
+        }
+      }
+    }
+  }
+
+  // Fake timestamps generated per element in collection.
+  Timestamp loop_internal_timestamp_ = Timestamp(0);
+};
+
+}  // namespace mediapipe
+
+#endif  // MEDIAPIPE_CALCULATORS_CORE_BEGIN_LOOP_CALCULATOR_H_

mediapipe/calculators/core/clip_detection_vector_size_calculator.cc

@@ -0,0 +1,26 @@
+// 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 <vector>
+
+#include "mediapipe/calculators/core/clip_vector_size_calculator.h"
+#include "mediapipe/framework/formats/detection.pb.h"
+
+namespace mediapipe {
+
+typedef ClipVectorSizeCalculator<::mediapipe::Detection>
+    ClipDetectionVectorSizeCalculator;
+REGISTER_CALCULATOR(ClipDetectionVectorSizeCalculator);
+
+}  // namespace mediapipe

mediapipe/calculators/core/clip_vector_size_calculator.cc

@@ -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.
+
+#include "mediapipe/calculators/core/clip_vector_size_calculator.h"
+
+#include <vector>
+
+#include "mediapipe/framework/calculator_framework.h"
+#include "mediapipe/framework/formats/rect.pb.h"
+
+namespace mediapipe {
+
+typedef ClipVectorSizeCalculator<::mediapipe::NormalizedRect>
+    ClipNormalizedRectVectorSizeCalculator;
+REGISTER_CALCULATOR(ClipNormalizedRectVectorSizeCalculator);
+
+}  // namespace mediapipe

mediapipe/calculators/core/clip_vector_size_calculator.h

@@ -0,0 +1,137 @@
+// 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.
+
+#ifndef MEDIAPIPE_CALCULATORS_CORE_CLIP_VECTOR_SIZE_CALCULATOR_H_
+#define MEDIAPIPE_CALCULATORS_CORE_CLIP_VECTOR_SIZE_CALCULATOR_H_
+
+#include <type_traits>
+#include <vector>
+
+#include "mediapipe/calculators/core/clip_vector_size_calculator.pb.h"
+#include "mediapipe/framework/calculator_framework.h"
+#include "mediapipe/framework/port/canonical_errors.h"
+#include "mediapipe/framework/port/ret_check.h"
+#include "mediapipe/framework/port/status.h"
+
+namespace mediapipe {
+
+// Clips the size of the input vector of type T to a specified max_vec_size.
+// In a graph it will be used as:
+// node {
+//   calculator: "ClipIntVectorSizeCalculator"
+//   input_stream: "input_vector"
+//   output_stream: "output_vector"
+//   options {
+//     [mediapipe.ClipIntVectorSizeCalculatorOptions.ext] {
+//       max_vec_size: 5
+//     }
+//   }
+// }
+template <typename T>
+class ClipVectorSizeCalculator : public CalculatorBase {
+ public:
+  static ::mediapipe::Status GetContract(CalculatorContract* cc) {
+    RET_CHECK(cc->Inputs().NumEntries() == 1);
+    RET_CHECK(cc->Outputs().NumEntries() == 1);
+
+    if (cc->Options<::mediapipe::ClipVectorSizeCalculatorOptions>()
+            .max_vec_size() < 1) {
+      return ::mediapipe::InternalError(
+          "max_vec_size should be greater than or equal to 1.");
+    }
+
+    cc->Inputs().Index(0).Set<std::vector<T>>();
+    cc->Outputs().Index(0).Set<std::vector<T>>();
+
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Open(CalculatorContext* cc) override {
+    cc->SetOffset(TimestampDiff(0));
+    max_vec_size_ = cc->Options<::mediapipe::ClipVectorSizeCalculatorOptions>()
+                        .max_vec_size();
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Process(CalculatorContext* cc) override {
+    if (max_vec_size_ < 1) {
+      return ::mediapipe::InternalError(
+          "max_vec_size should be greater than or equal to 1.");
+    }
+    if (cc->Inputs().Index(0).IsEmpty()) {
+      return ::mediapipe::OkStatus();
+    }
+
+    return ClipVectorSize<T>(std::is_copy_constructible<T>(), cc);
+  }
+
+  template <typename U>
+  ::mediapipe::Status ClipVectorSize(std::true_type, CalculatorContext* cc) {
+    auto output = absl::make_unique<std::vector<U>>();
+    const std::vector<U>& input_vector =
+        cc->Inputs().Index(0).Get<std::vector<U>>();
+    if (max_vec_size_ >= input_vector.size()) {
+      output->insert(output->end(), input_vector.begin(), input_vector.end());
+    } else {
+      for (int i = 0; i < max_vec_size_; ++i) {
+        output->push_back(input_vector[i]);
+      }
+    }
+    cc->Outputs().Index(0).Add(output.release(), cc->InputTimestamp());
+    return ::mediapipe::OkStatus();
+  }
+
+  template <typename U>
+  ::mediapipe::Status ClipVectorSize(std::false_type, CalculatorContext* cc) {
+    return ConsumeAndClipVectorSize<T>(std::is_move_constructible<U>(), cc);
+  }
+
+  template <typename U>
+  ::mediapipe::Status ConsumeAndClipVectorSize(std::true_type,
+                                               CalculatorContext* cc) {
+    auto output = absl::make_unique<std::vector<U>>();
+    ::mediapipe::StatusOr<std::unique_ptr<std::vector<U>>> input_status =
+        cc->Inputs().Index(0).Value().Consume<std::vector<U>>();
+
+    if (input_status.ok()) {
+      std::unique_ptr<std::vector<U>> input_vector =
+          std::move(input_status).ValueOrDie();
+      auto begin_it = input_vector->begin();
+      auto end_it = input_vector->end();
+      if (max_vec_size_ < input_vector->size()) {
+        end_it = input_vector->begin() + max_vec_size_;
+      }
+      output->insert(output->end(), std::make_move_iterator(begin_it),
+                     std::make_move_iterator(end_it));
+    } else {
+      return input_status.status();
+    }
+    cc->Outputs().Index(0).Add(output.release(), cc->InputTimestamp());
+    return ::mediapipe::OkStatus();
+  }
+
+  template <typename U>
+  ::mediapipe::Status ConsumeAndClipVectorSize(std::false_type,
+                                               CalculatorContext* cc) {
+    return ::mediapipe::InternalError(
+        "Cannot copy or move input vectors and clip their size.");
+  }
+
+ private:
+  int max_vec_size_ = 0;
+};
+
+}  // namespace mediapipe
+
+#endif  // MEDIAPIPE_CALCULATORS_CORE_CLIP_VECTOR_SIZE_CALCULATOR_H_

mediapipe/calculators/core/clip_vector_size_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 ClipVectorSizeCalculatorOptions {
+  extend CalculatorOptions {
+    optional ClipVectorSizeCalculatorOptions ext = 274674998;
+  }
+
+  // Maximum size of output vector.
+  optional int32 max_vec_size = 1 [default = 1];
+}

mediapipe/calculators/core/clip_vector_size_calculator_test.cc

@@ -0,0 +1,179 @@
+// 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/core/clip_vector_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 ClipVectorSizeCalculator<int> TestClipIntVectorSizeCalculator;
+REGISTER_CALCULATOR(TestClipIntVectorSizeCalculator);
+
+void AddInputVector(const std::vector<int>& input, int64 timestamp,
+                    CalculatorRunner* runner) {
+  runner->MutableInputs()->Index(0).packets.push_back(
+      MakePacket<std::vector<int>>(input).At(Timestamp(timestamp)));
+}
+
+TEST(TestClipIntVectorSizeCalculatorTest, EmptyVectorInput) {
+  CalculatorGraphConfig::Node node_config =
+      ParseTextProtoOrDie<CalculatorGraphConfig::Node>(R"(
+        calculator: "TestClipIntVectorSizeCalculator"
+        input_stream: "input_vector"
+        output_stream: "output_vector"
+        options {
+          [mediapipe.ClipVectorSizeCalculatorOptions.ext] { max_vec_size: 1 }
+        }
+      )");
+  CalculatorRunner runner(node_config);
+
+  std::vector<int> input = {};
+  AddInputVector(input, /*timestamp=*/1, &runner);
+  MP_ASSERT_OK(runner.Run());
+
+  const std::vector<Packet>& outputs = runner.Outputs().Index(0).packets;
+  EXPECT_EQ(1, outputs.size());
+  EXPECT_EQ(Timestamp(1), outputs[0].Timestamp());
+  EXPECT_TRUE(outputs[0].Get<std::vector<int>>().empty());
+}
+
+TEST(TestClipIntVectorSizeCalculatorTest, OneTimestamp) {
+  CalculatorGraphConfig::Node node_config =
+      ParseTextProtoOrDie<CalculatorGraphConfig::Node>(R"(
+        calculator: "TestClipIntVectorSizeCalculator"
+        input_stream: "input_vector"
+        output_stream: "output_vector"
+        options {
+          [mediapipe.ClipVectorSizeCalculatorOptions.ext] { max_vec_size: 2 }
+        }
+      )");
+  CalculatorRunner runner(node_config);
+
+  std::vector<int> input = {0, 1, 2, 3};
+  AddInputVector(input, /*timestamp=*/1, &runner);
+  MP_ASSERT_OK(runner.Run());
+
+  const std::vector<Packet>& outputs = runner.Outputs().Index(0).packets;
+  EXPECT_EQ(1, outputs.size());
+  EXPECT_EQ(Timestamp(1), outputs[0].Timestamp());
+  const std::vector<int>& output = outputs[0].Get<std::vector<int>>();
+  EXPECT_EQ(2, output.size());
+  std::vector<int> expected_vector = {0, 1};
+  EXPECT_EQ(expected_vector, output);
+}
+
+TEST(TestClipIntVectorSizeCalculatorTest, TwoInputsAtTwoTimestamps) {
+  CalculatorGraphConfig::Node node_config =
+      ParseTextProtoOrDie<CalculatorGraphConfig::Node>(R"(
+        calculator: "TestClipIntVectorSizeCalculator"
+        input_stream: "input_vector"
+        output_stream: "output_vector"
+        options {
+          [mediapipe.ClipVectorSizeCalculatorOptions.ext] { max_vec_size: 3 }
+        }
+      )");
+  CalculatorRunner runner(node_config);
+
+  {
+    std::vector<int> input = {0, 1, 2, 3};
+    AddInputVector(input, /*timestamp=*/1, &runner);
+  }
+  {
+    std::vector<int> input = {2, 3, 4, 5};
+    AddInputVector(input, /*timestamp=*/2, &runner);
+  }
+  MP_ASSERT_OK(runner.Run());
+
+  const std::vector<Packet>& outputs = runner.Outputs().Index(0).packets;
+  EXPECT_EQ(2, outputs.size());
+  {
+    EXPECT_EQ(Timestamp(1), outputs[0].Timestamp());
+    const std::vector<int>& output = outputs[0].Get<std::vector<int>>();
+    EXPECT_EQ(3, output.size());
+    std::vector<int> expected_vector = {0, 1, 2};
+    EXPECT_EQ(expected_vector, output);
+  }
+  {
+    EXPECT_EQ(Timestamp(2), outputs[1].Timestamp());
+    const std::vector<int>& output = outputs[1].Get<std::vector<int>>();
+    EXPECT_EQ(3, output.size());
+    std::vector<int> expected_vector = {2, 3, 4};
+    EXPECT_EQ(expected_vector, output);
+  }
+}
+
+typedef ClipVectorSizeCalculator<std::unique_ptr<int>>
+    TestClipUniqueIntPtrVectorSizeCalculator;
+REGISTER_CALCULATOR(TestClipUniqueIntPtrVectorSizeCalculator);
+
+TEST(TestClipUniqueIntPtrVectorSizeCalculatorTest, ConsumeOneTimestamp) {
+  /* Note: We don't use CalculatorRunner for this test because it keeps copies
+   * of input packets, so packets sent to the graph don't have sole ownership.
+   * The test needs to send packets that own the data.
+   */
+  CalculatorGraphConfig graph_config =
+      ParseTextProtoOrDie<CalculatorGraphConfig>(R"(
+        input_stream: "input_vector"
+        node {
+          calculator: "TestClipUniqueIntPtrVectorSizeCalculator"
+          input_stream: "input_vector"
+          output_stream: "output_vector"
+          options {
+            [mediapipe.ClipVectorSizeCalculatorOptions.ext] { max_vec_size: 3 }
+          }
+        }
+      )");
+
+  std::vector<Packet> outputs;
+  tool::AddVectorSink("output_vector", &graph_config, &outputs);
+
+  CalculatorGraph graph;
+  MP_EXPECT_OK(graph.Initialize(graph_config));
+  MP_EXPECT_OK(graph.StartRun({}));
+
+  // input1 : {0, 1, 2, 3, 4, 5}
+  auto input_vector = absl::make_unique<std::vector<std::unique_ptr<int>>>(6);
+  for (int i = 0; i < 6; ++i) {
+    input_vector->at(i) = absl::make_unique<int>(i);
+  }
+
+  MP_EXPECT_OK(graph.AddPacketToInputStream(
+      "input_vector", Adopt(input_vector.release()).At(Timestamp(1))));
+
+  MP_EXPECT_OK(graph.WaitUntilIdle());
+  MP_EXPECT_OK(graph.CloseAllPacketSources());
+  MP_EXPECT_OK(graph.WaitUntilDone());
+
+  EXPECT_EQ(1, outputs.size());
+  EXPECT_EQ(Timestamp(1), outputs[0].Timestamp());
+  const std::vector<std::unique_ptr<int>>& result =
+      outputs[0].Get<std::vector<std::unique_ptr<int>>>();
+  EXPECT_EQ(3, result.size());
+  for (int i = 0; i < 3; ++i) {
+    const std::unique_ptr<int>& v = result[i];
+    EXPECT_EQ(i, *v);
+  }
+}
+
+}  // namespace mediapipe

mediapipe/calculators/core/concatenate_vector_calculator.cc

@@ -19,7 +19,7 @@
 #include "mediapipe/framework/formats/landmark.pb.h"
 #include "tensorflow/lite/interpreter.h"
 
-#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
 #include "tensorflow/lite/delegates/gpu/gl/gl_buffer.h"
 #endif  //  !MEDIAPIPE_DISABLE_GPU
 
@@ -50,7 +50,7 @@ typedef ConcatenateVectorCalculator<::mediapipe::NormalizedLandmark>
     ConcatenateLandmarkVectorCalculator;
 REGISTER_CALCULATOR(ConcatenateLandmarkVectorCalculator);
 
-#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__APPLE__)
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
 typedef ConcatenateVectorCalculator<::tflite::gpu::gl::GlBuffer>
     ConcatenateGlBufferVectorCalculator;
 REGISTER_CALCULATOR(ConcatenateGlBufferVectorCalculator);

mediapipe/calculators/core/end_loop_calculator.cc

@@ -0,0 +1,40 @@
+// 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/core/end_loop_calculator.h"
+
+#include <vector>
+
+#include "mediapipe/framework/formats/landmark.pb.h"
+#include "mediapipe/framework/formats/rect.pb.h"
+#include "mediapipe/util/render_data.pb.h"
+
+namespace mediapipe {
+
+typedef EndLoopCalculator<std::vector<::mediapipe::NormalizedRect>>
+    EndLoopNormalizedRectCalculator;
+REGISTER_CALCULATOR(EndLoopNormalizedRectCalculator);
+
+typedef EndLoopCalculator<std::vector<::mediapipe::NormalizedLandmarkList>>
+    EndLoopNormalizedLandmarkListVectorCalculator;
+REGISTER_CALCULATOR(EndLoopNormalizedLandmarkListVectorCalculator);
+
+typedef EndLoopCalculator<std::vector<bool>> EndLoopBooleanCalculator;
+REGISTER_CALCULATOR(EndLoopBooleanCalculator);
+
+typedef EndLoopCalculator<std::vector<::mediapipe::RenderData>>
+    EndLoopRenderDataCalculator;
+REGISTER_CALCULATOR(EndLoopRenderDataCalculator);
+
+}  // namespace mediapipe

mediapipe/calculators/core/end_loop_calculator.h

@@ -0,0 +1,106 @@
+// 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.
+
+#ifndef MEDIAPIPE_CALCULATORS_CORE_END_LOOP_CALCULATOR_H_
+#define MEDIAPIPE_CALCULATORS_CORE_END_LOOP_CALCULATOR_H_
+
+#include "mediapipe/framework/calculator_context.h"
+#include "mediapipe/framework/calculator_contract.h"
+#include "mediapipe/framework/calculator_framework.h"
+#include "mediapipe/framework/collection_item_id.h"
+#include "mediapipe/framework/port/integral_types.h"
+#include "mediapipe/framework/port/ret_check.h"
+#include "mediapipe/framework/port/status.h"
+
+namespace mediapipe {
+
+// Calculator for completing the processing of loops on iterable collections
+// inside a MediaPipe graph. The EndLoopCalculator collects all input packets
+// from ITEM input_stream into a collection and upon receiving the flush signal
+// from the "BATCH_END" tagged input stream, it emits the aggregated results
+// at the original timestamp contained in the "BATCH_END" input stream.
+//
+// It is designed to be used like:
+//
+// node {
+//   calculator:    "BeginLoopWithIterableCalculator"
+//   input_stream:  "ITERABLE:input_iterable"      # IterableT @ext_ts
+//   output_stream: "ITEM:input_element"           # ItemT     @loop_internal_ts
+//   output_stream: "BATCH_END:ext_ts"             # Timestamp @loop_internal_ts
+// }
+//
+// node {
+//   calculator:    "ElementToBlaConverterSubgraph"
+//   input_stream:  "ITEM:input_to_loop_body"      # ItemT     @loop_internal_ts
+//   output_stream: "BLA:output_of_loop_body"      # ItemU     @loop_internal_ts
+// }
+//
+// node {
+//   calculator:    "EndLoopWithOutputCalculator"
+//   input_stream:  "ITEM:output_of_loop_body"     # ItemU     @loop_internal_ts
+//   input_stream:  "BATCH_END:ext_ts"             # Timestamp @loop_internal_ts
+//   output_stream: "OUTPUT:aggregated_result"     # IterableU @ext_ts
+// }
+template <typename IterableT>
+class EndLoopCalculator : public CalculatorBase {
+  using ItemT = typename IterableT::value_type;
+
+ public:
+  static ::mediapipe::Status GetContract(CalculatorContract* cc) {
+    RET_CHECK(cc->Inputs().HasTag("BATCH_END"))
+        << "Missing BATCH_END tagged input_stream.";
+    cc->Inputs().Tag("BATCH_END").Set<Timestamp>();
+
+    RET_CHECK(cc->Inputs().HasTag("ITEM"));
+    cc->Inputs().Tag("ITEM").Set<ItemT>();
+
+    RET_CHECK(cc->Outputs().HasTag("ITERABLE"));
+    cc->Outputs().Tag("ITERABLE").Set<IterableT>();
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Process(CalculatorContext* cc) override {
+    if (!cc->Inputs().Tag("ITEM").IsEmpty()) {
+      if (!input_stream_collection_) {
+        input_stream_collection_.reset(new IterableT);
+      }
+      input_stream_collection_->push_back(
+          cc->Inputs().Tag("ITEM").template Get<ItemT>());
+    }
+
+    if (!cc->Inputs().Tag("BATCH_END").Value().IsEmpty()) {  // flush signal
+      Timestamp loop_control_ts =
+          cc->Inputs().Tag("BATCH_END").template Get<Timestamp>();
+      if (input_stream_collection_) {
+        cc->Outputs()
+            .Tag("ITERABLE")
+            .Add(input_stream_collection_.release(), loop_control_ts);
+      } else {
+        // Since there is no collection, inform downstream calculators to not
+        // expect any packet by updating the timestamp bounds.
+        cc->Outputs()
+            .Tag("ITERABLE")
+            .SetNextTimestampBound(Timestamp(loop_control_ts.Value() + 1));
+      }
+    }
+    return ::mediapipe::OkStatus();
+  }
+
+ private:
+  std::unique_ptr<IterableT> input_stream_collection_;
+};
+
+}  // namespace mediapipe
+
+#endif  // MEDIAPIPE_CALCULATORS_CORE_END_LOOP_CALCULATOR_H_

mediapipe/calculators/core/packet_resampler_calculator.cc

@@ -74,6 +74,12 @@ class PacketResamplerCalculator : public CalculatorBase {
   ::mediapipe::Status Process(CalculatorContext* cc) override;
 
  private:
+  // Calculates the first sampled timestamp that incorporates a jittering
+  // offset.
+  void InitializeNextOutputTimestampWithJitter();
+  // Calculates the next sampled timestamp that incorporates a jittering offset.
+  void UpdateNextOutputTimestampWithJitter();
+
   // Logic for Process() when jitter_ != 0.0.
   ::mediapipe::Status ProcessWithJitter(CalculatorContext* cc);
 
@@ -233,6 +239,7 @@ TimestampDiff TimestampDiffFromSeconds(double seconds) {
       << Timestamp::kTimestampUnitsPerSecond;
 
   frame_time_usec_ = static_cast<int64>(1000000.0 / frame_rate_);
+
   video_header_.frame_rate = frame_rate_;
 
   if (resampler_options.output_header() !=
@@ -295,6 +302,17 @@ TimestampDiff TimestampDiffFromSeconds(double seconds) {
   return ::mediapipe::OkStatus();
 }
 
+void PacketResamplerCalculator::InitializeNextOutputTimestampWithJitter() {
+  next_output_timestamp_ =
+      first_timestamp_ + frame_time_usec_ * random_->RandFloat();
+}
+
+void PacketResamplerCalculator::UpdateNextOutputTimestampWithJitter() {
+  next_output_timestamp_ +=
+      frame_time_usec_ *
+      ((1.0 - jitter_) + 2.0 * jitter_ * random_->RandFloat());
+}
+
 ::mediapipe::Status PacketResamplerCalculator::ProcessWithJitter(
     CalculatorContext* cc) {
   RET_CHECK_GT(cc->InputTimestamp(), Timestamp::PreStream());
@@ -302,8 +320,13 @@ TimestampDiff TimestampDiffFromSeconds(double seconds) {
 
   if (first_timestamp_ == Timestamp::Unset()) {
     first_timestamp_ = cc->InputTimestamp();
-    next_output_timestamp_ =
+    InitializeNextOutputTimestampWithJitter();
-        first_timestamp_ + frame_time_usec_ * random_->RandFloat();
+    if (first_timestamp_ == next_output_timestamp_) {
+      OutputWithinLimits(
+          cc,
+          cc->Inputs().Get(input_data_id_).Value().At(next_output_timestamp_));
+      UpdateNextOutputTimestampWithJitter();
+    }
     return ::mediapipe::OkStatus();
   }
 
@@ -322,9 +345,7 @@ TimestampDiff TimestampDiffFromSeconds(double seconds) {
                               ? last_packet_
                               : cc->Inputs().Get(input_data_id_).Value())
                              .At(next_output_timestamp_));
-  next_output_timestamp_ +=
+  UpdateNextOutputTimestampWithJitter();
-      frame_time_usec_ *
-      ((1.0 - jitter_) + 2.0 * jitter_ * random_->RandFloat());
   return ::mediapipe::OkStatus();
 }
 

mediapipe/calculators/core/previous_loopback_calculator_test.cc

@@ -93,14 +93,19 @@ TEST(PreviousLoopbackCalculator, CorrectTimestamps) {
   EXPECT_EQ(TimestampValues(in_prev), (std::vector<int64>{1}));
   EXPECT_EQ(pair_values(in_prev.back()), std::make_pair(1, -1));
 
+  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));
+
   send_packet("in", 5);
   MP_EXPECT_OK(graph_.WaitUntilIdle());
-  EXPECT_EQ(TimestampValues(in_prev), (std::vector<int64>{1, 5}));
+  EXPECT_EQ(TimestampValues(in_prev), (std::vector<int64>{1, 2, 5}));
-  EXPECT_EQ(pair_values(in_prev.back()), std::make_pair(5, 1));
+  EXPECT_EQ(pair_values(in_prev.back()), std::make_pair(5, 2));
 
   send_packet("in", 15);
   MP_EXPECT_OK(graph_.WaitUntilIdle());
-  EXPECT_EQ(TimestampValues(in_prev), (std::vector<int64>{1, 5, 15}));
+  EXPECT_EQ(TimestampValues(in_prev), (std::vector<int64>{1, 2, 5, 15}));
   EXPECT_EQ(pair_values(in_prev.back()), std::make_pair(15, 5));
 
   MP_EXPECT_OK(graph_.CloseAllInputStreams());
@@ -182,18 +187,22 @@ TEST(PreviousLoopbackCalculator, ClosesCorrectly) {
   MP_EXPECT_OK(graph_.WaitUntilIdle());
   EXPECT_EQ(TimestampValues(outputs), (std::vector<int64>{1}));
 
+  send_packet("in", 2);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_EQ(TimestampValues(outputs), (std::vector<int64>{1, 2}));
+
   send_packet("in", 5);
   MP_EXPECT_OK(graph_.WaitUntilIdle());
-  EXPECT_EQ(TimestampValues(outputs), (std::vector<int64>{1, 5}));
+  EXPECT_EQ(TimestampValues(outputs), (std::vector<int64>{1, 2, 5}));
 
   send_packet("in", 15);
   MP_EXPECT_OK(graph_.WaitUntilIdle());
-  EXPECT_EQ(TimestampValues(outputs), (std::vector<int64>{1, 5, 15}));
+  EXPECT_EQ(TimestampValues(outputs), (std::vector<int64>{1, 2, 5, 15}));
 
   MP_EXPECT_OK(graph_.CloseAllInputStreams());
   MP_EXPECT_OK(graph_.WaitUntilIdle());
   EXPECT_EQ(TimestampValues(outputs),
-            (std::vector<int64>{1, 5, 15, Timestamp::Max().Value()}));
+            (std::vector<int64>{1, 2, 5, 15, Timestamp::Max().Value()}));
 
   MP_EXPECT_OK(graph_.WaitUntilDone());
 }

mediapipe/calculators/core/split_vector_calculator.cc

@@ -17,6 +17,7 @@
 #include <vector>
 
 #include "mediapipe/framework/formats/landmark.pb.h"
+#include "mediapipe/framework/formats/rect.pb.h"
 #include "tensorflow/lite/interpreter.h"
 
 namespace mediapipe {
@@ -41,4 +42,8 @@ REGISTER_CALCULATOR(SplitTfLiteTensorVectorCalculator);
 typedef SplitVectorCalculator<::mediapipe::NormalizedLandmark>
     SplitLandmarkVectorCalculator;
 REGISTER_CALCULATOR(SplitLandmarkVectorCalculator);
+
+typedef SplitVectorCalculator<::mediapipe::NormalizedRect>
+    SplitNormalizedRectVectorCalculator;
+REGISTER_CALCULATOR(SplitNormalizedRectVectorCalculator);
 }  // namespace mediapipe

mediapipe/calculators/core/stream_to_side_packet_calculator.cc

@@ -0,0 +1,48 @@
+// 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/framework/calculator_framework.h"
+#include "mediapipe/framework/port/status.h"
+#include "mediapipe/framework/timestamp.h"
+
+namespace mediapipe {
+
+// A calculator that takes a packet of an input stream and converts it to an
+// output side packet. This calculator only works under the assumption that the
+// input stream only has a single packet passing through.
+//
+// Example config:
+// node {
+//   calculator: "StreamToSidePacketCalculator"
+//   input_stream: "stream"
+//   output_side_packet: "side_packet"
+// }
+class StreamToSidePacketCalculator : public mediapipe::CalculatorBase {
+ public:
+  static mediapipe::Status GetContract(mediapipe::CalculatorContract* cc) {
+    cc->Inputs().Index(0).SetAny();
+    cc->OutputSidePackets().Index(0).SetAny();
+    return mediapipe::OkStatus();
+  }
+
+  mediapipe::Status Process(mediapipe::CalculatorContext* cc) override {
+    mediapipe::Packet& packet = cc->Inputs().Index(0).Value();
+    cc->OutputSidePackets().Index(0).Set(
+        packet.At(mediapipe::Timestamp::Unset()));
+    return mediapipe::OkStatus();
+  }
+};
+REGISTER_CALCULATOR(StreamToSidePacketCalculator);
+
+}  // namespace mediapipe

mediapipe/calculators/core/stream_to_side_packet_calculator_test.cc

@@ -0,0 +1,67 @@
+// 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 "absl/memory/memory.h"
+#include "mediapipe/framework/calculator_runner.h"
+#include "mediapipe/framework/packet.h"
+#include "mediapipe/framework/port/gmock.h"
+#include "mediapipe/framework/port/gtest.h"
+#include "mediapipe/framework/port/status.h"
+#include "mediapipe/framework/port/status_matchers.h"
+#include "mediapipe/framework/timestamp.h"
+
+namespace mediapipe {
+
+using ::testing::Test;
+
+class StreamToSidePacketCalculatorTest : public Test {
+ protected:
+  StreamToSidePacketCalculatorTest() {
+    const char kConfig[] = R"(
+      calculator: "StreamToSidePacketCalculator"
+      input_stream: "stream"
+      output_side_packet: "side_packet"
+    )";
+    runner_ = absl::make_unique<CalculatorRunner>(kConfig);
+  }
+
+  std::unique_ptr<CalculatorRunner> runner_;
+};
+
+TEST_F(StreamToSidePacketCalculatorTest,
+       StreamToSidePacketCalculatorWithEmptyStreamFails) {
+  EXPECT_EQ(runner_->Run().code(), mediapipe::StatusCode::kUnavailable);
+}
+
+TEST_F(StreamToSidePacketCalculatorTest,
+       StreamToSidePacketCalculatorWithSinglePacketCreatesSidePacket) {
+  runner_->MutableInputs()->Index(0).packets.push_back(
+      Adopt(new std::string("test")).At(Timestamp(1)));
+  MP_ASSERT_OK(runner_->Run());
+  EXPECT_EQ(runner_->OutputSidePackets().Index(0).Get<std::string>(), "test");
+}
+
+TEST_F(StreamToSidePacketCalculatorTest,
+       StreamToSidePacketCalculatorWithMultiplePacketsFails) {
+  runner_->MutableInputs()->Index(0).packets.push_back(
+      Adopt(new std::string("test1")).At(Timestamp(1)));
+  runner_->MutableInputs()->Index(0).packets.push_back(
+      Adopt(new std::string("test2")).At(Timestamp(2)));
+  EXPECT_EQ(runner_->Run().code(), mediapipe::StatusCode::kAlreadyExists);
+}
+
+}  // namespace mediapipe

mediapipe/calculators/image/BUILD

@@ -12,14 +12,14 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+load("//mediapipe/framework/port:build_config.bzl", "mediapipe_cc_proto_library")
+
 licenses(["notice"])  # Apache 2.0
 
 package(default_visibility = ["//visibility:private"])
 
 exports_files(["LICENSE"])
 
-load("//mediapipe/framework/port:build_config.bzl", "mediapipe_cc_proto_library")
-
 proto_library(
     name = "opencv_image_encoder_calculator_proto",
     srcs = ["opencv_image_encoder_calculator.proto"],
@@ -356,13 +356,13 @@ cc_library(
         "//mediapipe/framework/port:opencv_imgproc",
         "//mediapipe/framework/port:ret_check",
         "//mediapipe/framework/port:status",
-        "//mediapipe/gpu:gpu_buffer",
     ] + select({
         "//mediapipe/gpu:disable_gpu": [],
         "//conditions:default": [
             "//mediapipe/gpu:gl_calculator_helper",
             "//mediapipe/gpu:gl_simple_shaders",
             "//mediapipe/gpu:gl_quad_renderer",
+            "//mediapipe/gpu:gpu_buffer",
             "//mediapipe/gpu:shader_util",
         ],
     }),

mediapipe/calculators/image/image_transformation_calculator.cc

@@ -400,7 +400,7 @@ REGISTER_CALCULATOR(ImageTransformationCalculator);
   QuadRenderer* renderer = nullptr;
   GlTexture src1;
 
-#if defined(__APPLE__) && !TARGET_OS_OSX
+#if defined(MEDIAPIPE_IOS)
   if (input.format() == GpuBufferFormat::kBiPlanar420YpCbCr8VideoRange ||
       input.format() == GpuBufferFormat::kBiPlanar420YpCbCr8FullRange) {
     if (!yuv_renderer_) {

mediapipe/calculators/image/scale_image_calculator.proto

@@ -36,7 +36,8 @@ message ScaleImageCalculatorOptions {
 
   // If ratio is positive, crop the image to this minimum and maximum
   // aspect ratio (preserving the center of the frame).  This is done
-  // before scaling.
+  // before scaling.  The string must contain "/", so to disable cropping,
+  // set both to "0/1".
   //   For example, for a min_aspect_ratio of "9/16" and max of "16/9" the
   //   following cropping will occur:
   //       1920x1080 (which is 16:9) is not cropped

mediapipe/calculators/tensorflow/BUILD

@@ -13,12 +13,12 @@
 # limitations under the License.
 #
 
+load("//mediapipe/framework/port:build_config.bzl", "mediapipe_cc_proto_library")
+
 licenses(["notice"])  # Apache 2.0
 
 package(default_visibility = ["//visibility:private"])
 
-load("//mediapipe/framework/port:build_config.bzl", "mediapipe_cc_proto_library")
-
 proto_library(
     name = "graph_tensors_packet_generator_proto",
     srcs = ["graph_tensors_packet_generator.proto"],
@@ -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_cc",
+        "@org_tensorflow//tensorflow/core:protos_all",
     ],
     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_cc",
+        "@org_tensorflow//tensorflow/core:protos_all",
     ],
     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_cc",
+        "@org_tensorflow//tensorflow/core:protos_all",
     ],
     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_cc",
+        "@org_tensorflow//tensorflow/core:protos_all",
     ],
     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_cc",
+        "@org_tensorflow//tensorflow/core:protos_all",
     ],
     visibility = ["//visibility:public"],
     deps = [":vector_int_to_tensor_calculator_options_proto"],
@@ -296,7 +296,7 @@ cc_library(
     srcs = ["graph_tensors_packet_generator.cc"],
     visibility = ["//visibility:public"],
     deps = [
-        "//mediapipe/calculators/tensorflow:graph_tensors_packet_generator_cc_proto",
+        ":graph_tensors_packet_generator_cc_proto",
         "//mediapipe/framework:calculator_framework",
         "//mediapipe/framework/port:ret_check",
         "//mediapipe/framework/port:status",
@@ -311,7 +311,7 @@ cc_library(
     srcs = ["image_frame_to_tensor_calculator.cc"],
     visibility = ["//visibility:public"],
     deps = [
-        "//mediapipe/calculators/tensorflow:image_frame_to_tensor_calculator_cc_proto",
+        ":image_frame_to_tensor_calculator_cc_proto",
         "//mediapipe/framework:calculator_framework",
         "//mediapipe/framework/formats:image_frame",
         "//mediapipe/framework/port:ret_check",
@@ -333,7 +333,7 @@ cc_library(
     visibility = ["//visibility:public"],
     deps = [
         "//mediapipe/framework/formats:time_series_header_cc_proto",
-        "//mediapipe/calculators/tensorflow:matrix_to_tensor_calculator_options_cc_proto",
+        ":matrix_to_tensor_calculator_options_cc_proto",
         "//mediapipe/framework:calculator_framework",
         "//mediapipe/framework/formats:matrix",
         "//mediapipe/framework/port:status",
@@ -354,7 +354,7 @@ cc_library(
     srcs = ["lapped_tensor_buffer_calculator.cc"],
     visibility = ["//visibility:public"],
     deps = [
-        "//mediapipe/calculators/tensorflow:lapped_tensor_buffer_calculator_cc_proto",
+        ":lapped_tensor_buffer_calculator_cc_proto",
         "//mediapipe/framework:calculator_framework",
         "//mediapipe/framework/port:ret_check",
         "//mediapipe/framework/port:status",
@@ -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_cc",
+        "@org_tensorflow//tensorflow/core:protos_all",
     ],
     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_cc",
+        "@org_tensorflow//tensorflow/core:protos_all",
     ],
     alwayslink = 1,
 )
@@ -436,7 +436,7 @@ cc_library(
     visibility = ["//visibility:public"],
     deps = [
         ":tensorflow_session",
-        "//mediapipe/calculators/tensorflow:tensorflow_inference_calculator_cc_proto",
+        ":tensorflow_inference_calculator_cc_proto",
         "//mediapipe/framework:calculator_framework",
         "//mediapipe/framework/tool:status_util",
         "@com_google_absl//absl/strings",
@@ -514,7 +514,7 @@ cc_library(
     visibility = ["//visibility:public"],
     deps = [
         ":tensorflow_session",
-        "//mediapipe/calculators/tensorflow:tensorflow_session_from_frozen_graph_generator_cc_proto",
+        ":tensorflow_session_from_frozen_graph_generator_cc_proto",
         "//mediapipe/framework:calculator_framework",
         "//mediapipe/framework/tool:status_util",
         "//mediapipe/framework/port:status",
@@ -573,7 +573,7 @@ cc_library(
     visibility = ["//visibility:public"],
     deps = [
         ":tensorflow_session",
-        "//mediapipe/calculators/tensorflow:tensorflow_session_from_saved_model_generator_cc_proto",
+        ":tensorflow_session_from_saved_model_generator_cc_proto",
         "//mediapipe/framework:packet_generator",
         "//mediapipe/framework:packet_type",
         "//mediapipe/framework/tool:status_util",
@@ -597,7 +597,7 @@ cc_library(
     srcs = ["tensor_squeeze_dimensions_calculator.cc"],
     visibility = ["//visibility:public"],
     deps = [
-        "//mediapipe/calculators/tensorflow:tensor_squeeze_dimensions_calculator_cc_proto",
+        ":tensor_squeeze_dimensions_calculator_cc_proto",
         "//mediapipe/framework:calculator_framework",
         "//mediapipe/framework/port:ret_check",
         "//mediapipe/framework/port:status",
@@ -611,7 +611,7 @@ cc_library(
     srcs = ["tensor_to_image_frame_calculator.cc"],
     visibility = ["//visibility:public"],
     deps = [
-        "//mediapipe/calculators/tensorflow:tensor_to_image_frame_calculator_cc_proto",
+        ":tensor_to_image_frame_calculator_cc_proto",
         "//mediapipe/framework:calculator_framework",
         "//mediapipe/framework/formats:image_frame",
         "//mediapipe/framework/port:ret_check",
@@ -627,7 +627,7 @@ cc_library(
     visibility = ["//visibility:public"],
     deps = [
         "//mediapipe/framework/formats:time_series_header_cc_proto",
-        "//mediapipe/calculators/tensorflow:tensor_to_matrix_calculator_cc_proto",
+        ":tensor_to_matrix_calculator_cc_proto",
         "//mediapipe/framework:calculator_framework",
         "//mediapipe/framework/formats:matrix",
         "//mediapipe/framework/port:status",
@@ -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_cc",
+        "@org_tensorflow//tensorflow/core:protos_all",
     ],
     alwayslink = 1,
 )
@@ -667,7 +667,7 @@ cc_library(
         "//mediapipe/framework:calculator_framework",
         "//mediapipe/framework/port:status",
         "//mediapipe/framework/port:ret_check",
-        "//mediapipe/calculators/tensorflow:tensor_to_vector_float_calculator_options_cc_proto",
+        ":tensor_to_vector_float_calculator_options_cc_proto",
     ] + select({
         "//conditions:default": [
             "@org_tensorflow//tensorflow/core:framework",
@@ -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_cc",
+        "@org_tensorflow//tensorflow/core:protos_all",
     ],
     alwayslink = 1,
 )
@@ -719,7 +719,7 @@ cc_library(
     srcs = ["vector_float_to_tensor_calculator.cc"],
     visibility = ["//visibility:public"],
     deps = [
-        "//mediapipe/calculators/tensorflow:vector_float_to_tensor_calculator_options_cc_proto",
+        ":vector_float_to_tensor_calculator_options_cc_proto",
         "//mediapipe/framework:calculator_framework",
         "//mediapipe/framework/port:ret_check",
         "//mediapipe/framework/port:status",
@@ -733,11 +733,11 @@ cc_library(
     srcs = ["unpack_yt8m_sequence_example_calculator.cc"],
     visibility = ["//visibility:public"],
     deps = [
-        "//mediapipe/calculators/tensorflow:lapped_tensor_buffer_calculator_cc_proto",
+        ":lapped_tensor_buffer_calculator_cc_proto",
         "//mediapipe/framework:calculator_framework",
         "//mediapipe/framework:packet",
         "//mediapipe/framework/port:status",
-        "@org_tensorflow//tensorflow/core:protos_all_cc",
+        "@org_tensorflow//tensorflow/core:protos_all",
     ],
     alwayslink = 1,
 )
@@ -747,7 +747,7 @@ cc_test(
     srcs = ["graph_tensors_packet_generator_test.cc"],
     deps = [
         ":graph_tensors_packet_generator",
-        "//mediapipe/calculators/tensorflow:graph_tensors_packet_generator_cc_proto",
+        ":graph_tensors_packet_generator_cc_proto",
         "//mediapipe/framework:packet",
         "//mediapipe/framework:packet_generator_cc_proto",
         "//mediapipe/framework:packet_set",
@@ -779,7 +779,7 @@ cc_test(
     srcs = ["matrix_to_tensor_calculator_test.cc"],
     deps = [
         ":matrix_to_tensor_calculator",
-        "//mediapipe/calculators/tensorflow:matrix_to_tensor_calculator_options_cc_proto",
+        ":matrix_to_tensor_calculator_options_cc_proto",
         "//mediapipe/framework:calculator_framework",
         "//mediapipe/framework:calculator_runner",
         "//mediapipe/framework/formats:matrix",
@@ -795,13 +795,13 @@ cc_test(
     srcs = ["lapped_tensor_buffer_calculator_test.cc"],
     deps = [
         ":lapped_tensor_buffer_calculator",
-        "//mediapipe/calculators/tensorflow:lapped_tensor_buffer_calculator_cc_proto",
+        ":lapped_tensor_buffer_calculator_cc_proto",
         "//mediapipe/framework:calculator_framework",
         "//mediapipe/framework:calculator_runner",
         "//mediapipe/framework/port:gtest_main",
         "@com_google_absl//absl/memory",
         "@org_tensorflow//tensorflow/core:framework",
-        "@org_tensorflow//tensorflow/core:protos_all_cc",
+        "@org_tensorflow//tensorflow/core:protos_all",
     ],
 )
 
@@ -840,7 +840,7 @@ cc_test(
         "//mediapipe/util/sequence:media_sequence",
         "@com_google_absl//absl/memory",
         "@com_google_absl//absl/strings",
-        "@org_tensorflow//tensorflow/core:protos_all_cc",
+        "@org_tensorflow//tensorflow/core:protos_all",
     ],
 )
 
@@ -867,7 +867,7 @@ cc_test(
         "@com_google_absl//absl/strings",
         "@org_tensorflow//tensorflow/core:direct_session",
         "@org_tensorflow//tensorflow/core:framework",
-        "@org_tensorflow//tensorflow/core:protos_all_cc",
+        "@org_tensorflow//tensorflow/core:protos_all",
         "@org_tensorflow//tensorflow/core:testlib",
         "@org_tensorflow//tensorflow/core/kernels:conv_ops",
         "@org_tensorflow//tensorflow/core/kernels:math",
@@ -883,7 +883,7 @@ cc_test(
         ":tensorflow_inference_calculator",
         ":tensorflow_session",
         ":tensorflow_session_from_frozen_graph_generator",
-        "//mediapipe/calculators/tensorflow:tensorflow_session_from_frozen_graph_generator_cc_proto",
+        ":tensorflow_session_from_frozen_graph_generator_cc_proto",
         "//mediapipe/framework:calculator_framework",
         "//mediapipe/framework:packet",
         "//mediapipe/framework:packet_generator_cc_proto",
@@ -897,7 +897,7 @@ cc_test(
         "@com_google_absl//absl/strings",
         "@org_tensorflow//tensorflow/core:direct_session",
         "@org_tensorflow//tensorflow/core:framework",
-        "@org_tensorflow//tensorflow/core:protos_all_cc",
+        "@org_tensorflow//tensorflow/core:protos_all",
         "@org_tensorflow//tensorflow/core:testlib",
         "@org_tensorflow//tensorflow/core/kernels:conv_ops",
         "@org_tensorflow//tensorflow/core/kernels:math",
@@ -913,7 +913,7 @@ cc_test(
         ":tensorflow_inference_calculator",
         ":tensorflow_session",
         ":tensorflow_session_from_saved_model_generator",
-        "//mediapipe/calculators/tensorflow:tensorflow_session_from_saved_model_generator_cc_proto",
+        ":tensorflow_session_from_saved_model_generator_cc_proto",
         "//mediapipe/framework:calculator_framework",
         "//mediapipe/framework:packet",
         "//mediapipe/framework:packet_generator_cc_proto",
@@ -923,14 +923,8 @@ cc_test(
         "//mediapipe/framework/tool:tag_map_helper",
         "//mediapipe/framework/tool:validate_type",
         "@com_google_absl//absl/strings",
+        "@org_tensorflow//tensorflow/core:all_kernels",
         "@org_tensorflow//tensorflow/core:direct_session",
-        "@org_tensorflow//tensorflow/core/kernels:array",
-        "@org_tensorflow//tensorflow/core/kernels:bitcast_op",
-        "@org_tensorflow//tensorflow/core/kernels:conv_ops",
-        "@org_tensorflow//tensorflow/core/kernels:io",
-        "@org_tensorflow//tensorflow/core/kernels:state",
-        "@org_tensorflow//tensorflow/core/kernels:string",
-        "@org_tensorflow//tensorflow/core/kernels/data:tensor_dataset_op",
     ],
 )
 
@@ -954,14 +948,8 @@ cc_test(
         "//mediapipe/framework/tool:tag_map_helper",
         "//mediapipe/framework/tool:validate_type",
         "@com_google_absl//absl/strings",
+        "@org_tensorflow//tensorflow/core:all_kernels",
         "@org_tensorflow//tensorflow/core:direct_session",
-        "@org_tensorflow//tensorflow/core/kernels:array",
-        "@org_tensorflow//tensorflow/core/kernels:bitcast_op",
-        "@org_tensorflow//tensorflow/core/kernels:conv_ops",
-        "@org_tensorflow//tensorflow/core/kernels:io",
-        "@org_tensorflow//tensorflow/core/kernels:state",
-        "@org_tensorflow//tensorflow/core/kernels:string",
-        "@org_tensorflow//tensorflow/core/kernels/data:tensor_dataset_op",
     ],
 )
 
@@ -970,12 +958,12 @@ cc_test(
     srcs = ["tensor_squeeze_dimensions_calculator_test.cc"],
     deps = [
         ":tensor_squeeze_dimensions_calculator",
-        "//mediapipe/calculators/tensorflow:tensor_squeeze_dimensions_calculator_cc_proto",
+        ":tensor_squeeze_dimensions_calculator_cc_proto",
         "//mediapipe/framework:calculator_framework",
         "//mediapipe/framework:calculator_runner",
         "//mediapipe/framework/port:gtest_main",
         "@org_tensorflow//tensorflow/core:framework",
-        "@org_tensorflow//tensorflow/core:protos_all_cc",
+        "@org_tensorflow//tensorflow/core:protos_all",
     ],
 )
 
@@ -985,13 +973,13 @@ cc_test(
     srcs = ["tensor_to_image_frame_calculator_test.cc"],
     deps = [
         ":tensor_to_image_frame_calculator",
-        "//mediapipe/calculators/tensorflow:tensor_to_image_frame_calculator_cc_proto",
+        ":tensor_to_image_frame_calculator_cc_proto",
         "//mediapipe/framework:calculator_framework",
         "//mediapipe/framework:calculator_runner",
         "//mediapipe/framework/formats:image_frame",
         "//mediapipe/framework/port:gtest_main",
         "@org_tensorflow//tensorflow/core:framework",
-        "@org_tensorflow//tensorflow/core:protos_all_cc",
+        "@org_tensorflow//tensorflow/core:protos_all",
     ],
 )
 
@@ -1001,14 +989,14 @@ cc_test(
     srcs = ["tensor_to_matrix_calculator_test.cc"],
     deps = [
         ":tensor_to_matrix_calculator",
-        "//mediapipe/calculators/tensorflow:tensor_to_matrix_calculator_cc_proto",
+        ":tensor_to_matrix_calculator_cc_proto",
         "//mediapipe/framework:calculator_framework",
         "//mediapipe/framework:calculator_runner",
         "//mediapipe/framework/formats:matrix",
         "//mediapipe/framework/formats:time_series_header_cc_proto",
         "//mediapipe/framework/port:gtest_main",
         "@org_tensorflow//tensorflow/core:framework",
-        "@org_tensorflow//tensorflow/core:protos_all_cc",
+        "@org_tensorflow//tensorflow/core:protos_all",
     ],
 )
 
@@ -1017,12 +1005,12 @@ cc_test(
     srcs = ["tensor_to_vector_float_calculator_test.cc"],
     deps = [
         ":tensor_to_vector_float_calculator",
-        "//mediapipe/calculators/tensorflow:tensor_to_vector_float_calculator_options_cc_proto",
+        ":tensor_to_vector_float_calculator_options_cc_proto",
         "//mediapipe/framework:calculator_framework",
         "//mediapipe/framework:calculator_runner",
         "//mediapipe/framework/port:gtest_main",
         "@org_tensorflow//tensorflow/core:framework",
-        "@org_tensorflow//tensorflow/core:protos_all_cc",
+        "@org_tensorflow//tensorflow/core:protos_all",
     ],
 )
 
@@ -1042,7 +1030,7 @@ cc_test(
         "//mediapipe/util/sequence:media_sequence",
         "@com_google_absl//absl/memory",
         "@com_google_absl//absl/strings",
-        "@org_tensorflow//tensorflow/core:protos_all_cc",
+        "@org_tensorflow//tensorflow/core:protos_all",
     ],
 )
 
@@ -1056,7 +1044,7 @@ cc_test(
         "//mediapipe/framework:calculator_runner",
         "//mediapipe/framework/port:gtest_main",
         "@org_tensorflow//tensorflow/core:framework",
-        "@org_tensorflow//tensorflow/core:protos_all_cc",
+        "@org_tensorflow//tensorflow/core:protos_all",
     ],
 )
 
@@ -1065,12 +1053,12 @@ cc_test(
     srcs = ["vector_float_to_tensor_calculator_test.cc"],
     deps = [
         ":vector_float_to_tensor_calculator",
-        "//mediapipe/calculators/tensorflow:vector_float_to_tensor_calculator_options_cc_proto",
+        ":vector_float_to_tensor_calculator_options_cc_proto",
         "//mediapipe/framework:calculator_framework",
         "//mediapipe/framework:calculator_runner",
         "//mediapipe/framework/port:gtest_main",
         "@org_tensorflow//tensorflow/core:framework",
-        "@org_tensorflow//tensorflow/core:protos_all_cc",
+        "@org_tensorflow//tensorflow/core:protos_all",
     ],
 )
 
@@ -1094,7 +1082,7 @@ cc_test(
         ":tensorflow_session",
         ":tensorflow_inference_calculator",
         ":tensorflow_session_from_frozen_graph_generator",
-        "//mediapipe/calculators/tensorflow:tensorflow_session_from_frozen_graph_generator_cc_proto",
+        ":tensorflow_session_from_frozen_graph_generator_cc_proto",
         "//mediapipe/framework/deps:file_path",
         "//mediapipe/framework:calculator_framework",
         "//mediapipe/framework:calculator_runner",

mediapipe/calculators/tensorflow/tensorflow_inference_calculator.cc

@@ -34,7 +34,7 @@
 #include "tensorflow/core/framework/tensor_shape.h"
 #include "tensorflow/core/framework/tensor_util.h"
 
-#if !defined(__ANDROID__) && !defined(__APPLE__)
+#if !defined(MEDIAPIPE_MOBILE) && !defined(__APPLE__)
 #include "tensorflow/core/profiler/lib/traceme.h"
 #endif
 
@@ -441,7 +441,7 @@ class TensorFlowInferenceCalculator : public CalculatorBase {
     const int64 run_start_time = absl::ToUnixMicros(clock_->TimeNow());
     tf::Status tf_status;
     {
-#if !defined(__ANDROID__) && !defined(__APPLE__)
+#if !defined(MEDIAPIPE_MOBILE) && !defined(__APPLE__)
       tensorflow::profiler::TraceMe trace(absl::string_view(cc->NodeName()));
 #endif
       tf_status = session_->Run(input_tensors, output_tensor_names,

mediapipe/calculators/tensorflow/tensorflow_session_from_frozen_graph_calculator.cc

@@ -31,8 +31,7 @@
 #include "mediapipe/framework/tool/status_util.h"
 #include "tensorflow/core/public/session_options.h"
 
-#if defined(MEDIAPIPE_LITE) || defined(__ANDROID__) || \
+#if defined(MEDIAPIPE_MOBILE)
-    defined(__APPLE__) && !TARGET_OS_OSX
 #include "mediapipe/util/android/file/base/helpers.h"
 #else
 #include "mediapipe/framework/port/file_helpers.h"

mediapipe/calculators/tensorflow/tfrecord_reader_calculator.cc

@@ -85,7 +85,7 @@ class TFRecordReaderCalculator : public CalculatorBase {
   tensorflow::io::RecordReader reader(file.get(),
                                       tensorflow::io::RecordReaderOptions());
   tensorflow::uint64 offset = 0;
-  std::string example_str;
+  tensorflow::tstring example_str;
   const int target_idx =
       cc->InputSidePackets().HasTag(kRecordIndex)
           ? cc->InputSidePackets().Tag(kRecordIndex).Get<int>()
@@ -98,7 +98,7 @@ class TFRecordReaderCalculator : public CalculatorBase {
     if (current_idx == target_idx) {
       if (cc->OutputSidePackets().HasTag(kExampleTag)) {
         tensorflow::Example tf_example;
-        tf_example.ParseFromString(example_str);
+        tf_example.ParseFromArray(example_str.data(), example_str.size());
         cc->OutputSidePackets()
             .Tag(kExampleTag)
             .Set(MakePacket<tensorflow::Example>(std::move(tf_example)));

mediapipe/calculators/tflite/BUILD

@@ -13,12 +13,12 @@
 # limitations under the License.
 #
 
+load("//mediapipe/framework/port:build_config.bzl", "mediapipe_cc_proto_library")
+
 licenses(["notice"])  # Apache 2.0
 
 package(default_visibility = ["//visibility:private"])
 
-load("//mediapipe/framework/port:build_config.bzl", "mediapipe_cc_proto_library")
-
 proto_library(
     name = "ssd_anchors_calculator_proto",
     srcs = ["ssd_anchors_calculator.proto"],
@@ -238,6 +238,7 @@ cc_library(
             "@org_tensorflow//tensorflow/lite/delegates/gpu/common:shape",
             "@org_tensorflow//tensorflow/lite/delegates/gpu/metal:buffer_convert",
             "@org_tensorflow//tensorflow/lite/delegates/gpu:metal_delegate",
+            "@org_tensorflow//tensorflow/lite/delegates/gpu:metal_delegate_internal",
         ],
         "//conditions:default": [
             "//mediapipe/gpu:gl_calculator_helper",
@@ -248,6 +249,11 @@ cc_library(
             "@org_tensorflow//tensorflow/lite/delegates/gpu/gl:gl_program",
             "@org_tensorflow//tensorflow/lite/delegates/gpu/gl:gl_shader",
         ],
+    }) + select({
+        "//conditions:default": [],
+        "//mediapipe:android": [
+            "@org_tensorflow//tensorflow/lite/delegates/nnapi:nnapi_delegate",
+        ],
     }),
     alwayslink = 1,
 )

mediapipe/calculators/tflite/tflite_converter_calculator.cc

@@ -25,8 +25,7 @@
 #include "tensorflow/lite/error_reporter.h"
 #include "tensorflow/lite/interpreter.h"
 
-#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__) && \
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-    !defined(__APPLE__)
 #include "mediapipe/gpu/gl_calculator_helper.h"
 #include "mediapipe/gpu/gpu_buffer.h"
 #include "tensorflow/lite/delegates/gpu/gl/gl_buffer.h"
@@ -35,7 +34,7 @@
 #include "tensorflow/lite/delegates/gpu/gl_delegate.h"
 #endif  //  !MEDIAPIPE_DISABLE_GPU
 
-#if defined(__APPLE__) && !TARGET_OS_OSX  // iOS
+#if defined(MEDIAPIPE_IOS)
 #import <CoreVideo/CoreVideo.h>
 #import <Metal/Metal.h>
 #import <MetalKit/MetalKit.h>
@@ -46,10 +45,9 @@
 #include "tensorflow/lite/delegates/gpu/metal_delegate.h"
 #endif  // iOS
 
-#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__) && \
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-    !defined(__APPLE__)
 typedef ::tflite::gpu::gl::GlBuffer GpuTensor;
-#elif defined(__APPLE__) && !TARGET_OS_OSX  // iOS
+#elif defined(MEDIAPIPE_IOS)
 typedef id<MTLBuffer> GpuTensor;
 #endif
 
@@ -69,8 +67,7 @@ typedef Eigen::Matrix<float, Eigen::Dynamic, Eigen::Dynamic, Eigen::ColMajor>
 
 namespace mediapipe {
 
-#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__) && \
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-    !defined(__APPLE__)
 using ::tflite::gpu::gl::CreateReadWriteShaderStorageBuffer;
 using ::tflite::gpu::gl::GlProgram;
 using ::tflite::gpu::gl::GlShader;
@@ -80,7 +77,7 @@ struct GPUData {
   GlShader shader;
   GlProgram program;
 };
-#elif defined(__APPLE__) && !TARGET_OS_OSX  // iOS
+#elif defined(MEDIAPIPE_IOS)
 struct GPUData {
   int elements = 1;
   GpuTensor buffer;
@@ -149,11 +146,10 @@ class TfLiteConverterCalculator : public CalculatorBase {
 
   std::unique_ptr<tflite::Interpreter> interpreter_ = nullptr;
 
-#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__) && \
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-    !defined(__APPLE__)
   mediapipe::GlCalculatorHelper gpu_helper_;
   std::unique_ptr<GPUData> gpu_data_out_;
-#elif defined(__APPLE__) && !TARGET_OS_OSX  // iOS
+#elif defined(MEDIAPIPE_IOS)
   MPPMetalHelper* gpu_helper_ = nullptr;
   std::unique_ptr<GPUData> gpu_data_out_;
 #endif
@@ -202,10 +198,9 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);
 #endif  //  !MEDIAPIPE_DISABLE_GPU
 
   if (use_gpu) {
-#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__) && \
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-    !defined(__APPLE__)
     MP_RETURN_IF_ERROR(mediapipe::GlCalculatorHelper::UpdateContract(cc));
-#elif defined(__APPLE__) && !TARGET_OS_OSX  // iOS
+#elif defined(MEDIAPIPE_IOS)
     MP_RETURN_IF_ERROR([MPPMetalHelper updateContract:cc]);
 #endif
   }
@@ -236,10 +231,9 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);
               cc->Outputs().HasTag("TENSORS_GPU"));
     // Cannot use quantization.
     use_quantized_tensors_ = false;
-#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__) && \
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-    !defined(__APPLE__)
     MP_RETURN_IF_ERROR(gpu_helper_.Open(cc));
-#elif defined(__APPLE__) && !TARGET_OS_OSX  // iOS
+#elif defined(MEDIAPIPE_IOS)
     gpu_helper_ = [[MPPMetalHelper alloc] initWithCalculatorContext:cc];
     RET_CHECK(gpu_helper_);
 #endif
@@ -270,11 +264,10 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);
 }
 
 ::mediapipe::Status TfLiteConverterCalculator::Close(CalculatorContext* cc) {
-#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__) && \
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-    !defined(__APPLE__)
   gpu_helper_.RunInGlContext([this] { gpu_data_out_.reset(); });
 #endif
-#if defined(__APPLE__) && !TARGET_OS_OSX  // iOS
+#if defined(MEDIAPIPE_IOS)
   gpu_data_out_.reset();
 #endif
   return ::mediapipe::OkStatus();
@@ -390,8 +383,7 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);
 
 ::mediapipe::Status TfLiteConverterCalculator::ProcessGPU(
     CalculatorContext* cc) {
-#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__) && \
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-    !defined(__APPLE__)
   // GpuBuffer to tflite::gpu::GlBuffer conversion.
   const auto& input = cc->Inputs().Tag("IMAGE_GPU").Get<mediapipe::GpuBuffer>();
   MP_RETURN_IF_ERROR(
@@ -427,7 +419,7 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);
   cc->Outputs()
       .Tag("TENSORS_GPU")
       .Add(output_tensors.release(), cc->InputTimestamp());
-#elif defined(__APPLE__) && !TARGET_OS_OSX  // iOS
+#elif defined(MEDIAPIPE_IOS)
   // GpuBuffer to id<MTLBuffer> conversion.
   const auto& input = cc->Inputs().Tag("IMAGE_GPU").Get<mediapipe::GpuBuffer>();
   {
@@ -493,8 +485,7 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);
     RET_CHECK_FAIL() << "Num input channels is less than desired output.";
 #endif  //  !MEDIAPIPE_DISABLE_GPU
 
-#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__) && \
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-    !defined(__APPLE__)
   MP_RETURN_IF_ERROR(gpu_helper_.RunInGlContext(
       [this, &include_alpha, &input, &single_channel]() -> ::mediapipe::Status {
         // Device memory.
@@ -538,7 +529,9 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);
                                                    &gpu_data_out_->program));
         return ::mediapipe::OkStatus();
       }));
-#elif defined(__APPLE__) && !TARGET_OS_OSX  // iOS
+
+#elif defined(MEDIAPIPE_IOS)
+
   RET_CHECK(include_alpha)
       << "iOS GPU inference currently accepts only RGBA input.";
 
@@ -619,7 +612,7 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);
   CHECK_GE(max_num_channels_, 1);
   CHECK_LE(max_num_channels_, 4);
   CHECK_NE(max_num_channels_, 2);
-#if defined(__APPLE__) && !TARGET_OS_OSX  // iOS
+#if defined(MEDIAPIPE_IOS)
   if (cc->Inputs().HasTag("IMAGE_GPU"))
     // Currently on iOS, tflite gpu input tensor must be 4 channels,
     // so input image must be 4 channels also (checked in InitGpu).

mediapipe/calculators/tflite/tflite_inference_calculator.cc

@@ -27,8 +27,7 @@
 #include "tensorflow/lite/kernels/register.h"
 #include "tensorflow/lite/model.h"
 
-#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__) && \
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-    !defined(__APPLE__)
 #include "mediapipe/gpu/gl_calculator_helper.h"
 #include "mediapipe/gpu/gpu_buffer.h"
 #include "tensorflow/lite/delegates/gpu/common/shape.h"
@@ -36,9 +35,9 @@
 #include "tensorflow/lite/delegates/gpu/gl/gl_program.h"
 #include "tensorflow/lite/delegates/gpu/gl/gl_shader.h"
 #include "tensorflow/lite/delegates/gpu/gl_delegate.h"
-#endif  //  !MEDIAPIPE_DISABLE_GPU
+#endif  //  !MEDIAPIPE_DISABLE_GL_COMPUTE
 
-#if defined(__APPLE__) && !TARGET_OS_OSX  // iOS
+#if defined(MEDIAPIPE_IOS)
 #import <CoreVideo/CoreVideo.h>
 #import <Metal/Metal.h>
 #import <MetalKit/MetalKit.h>
@@ -49,14 +48,18 @@
 #include "tensorflow/lite/delegates/gpu/common/shape.h"
 #include "tensorflow/lite/delegates/gpu/metal/buffer_convert.h"
 #include "tensorflow/lite/delegates/gpu/metal_delegate.h"
+#include "tensorflow/lite/delegates/gpu/metal_delegate_internal.h"
 #endif  // iOS
 
+#if defined(MEDIAPIPE_ANDROID)
+#include "tensorflow/lite/delegates/nnapi/nnapi_delegate.h"
+#endif  // ANDROID
+
 namespace {
 
-#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__) && \
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-    !defined(__APPLE__)
 typedef ::tflite::gpu::gl::GlBuffer GpuTensor;
-#elif defined(__APPLE__) && !TARGET_OS_OSX  // iOS
+#elif defined(MEDIAPIPE_IOS)
 typedef id<MTLBuffer> GpuTensor;
 #endif
 
@@ -64,14 +67,35 @@ typedef id<MTLBuffer> GpuTensor;
 size_t RoundUp(size_t n, size_t m) { return ((n + m - 1) / m) * m; }  // NOLINT
 }  // namespace
 
+#if defined(MEDIAPIPE_EDGE_TPU)
+#include "edgetpu.h"
+
+// Creates and returns an Edge TPU interpreter to run the given edgetpu model.
+std::unique_ptr<tflite::Interpreter> BuildEdgeTpuInterpreter(
+    const tflite::FlatBufferModel& model,
+    tflite::ops::builtin::BuiltinOpResolver* resolver,
+    edgetpu::EdgeTpuContext* edgetpu_context) {
+  resolver->AddCustom(edgetpu::kCustomOp, edgetpu::RegisterCustomOp());
+  std::unique_ptr<tflite::Interpreter> interpreter;
+  if (tflite::InterpreterBuilder(model, *resolver)(&interpreter) != kTfLiteOk) {
+    std::cerr << "Failed to build edge TPU interpreter." << std::endl;
+  }
+  interpreter->SetExternalContext(kTfLiteEdgeTpuContext, edgetpu_context);
+  interpreter->SetNumThreads(1);
+  if (interpreter->AllocateTensors() != kTfLiteOk) {
+    std::cerr << "Failed to allocate edge TPU tensors." << std::endl;
+  }
+  return interpreter;
+}
+#endif  // MEDIAPIPE_EDGE_TPU
+
 // TfLiteInferenceCalculator File Layout:
 //  * Header
 //  * Core
 //  * Aux
 namespace mediapipe {
 
-#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__) && \
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-    !defined(__APPLE__)
 using ::tflite::gpu::gl::CopyBuffer;
 using ::tflite::gpu::gl::CreateReadWriteShaderStorageBuffer;
 using ::tflite::gpu::gl::GlBuffer;
@@ -150,18 +174,22 @@ class TfLiteInferenceCalculator : public CalculatorBase {
   std::unique_ptr<tflite::FlatBufferModel> model_;
   TfLiteDelegate* delegate_ = nullptr;
 
-#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__) && \
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-    !defined(__APPLE__)
   mediapipe::GlCalculatorHelper gpu_helper_;
   std::unique_ptr<GPUData> gpu_data_in_;
   std::vector<std::unique_ptr<GPUData>> gpu_data_out_;
-#elif defined(__APPLE__) && !TARGET_OS_OSX  // iOS
+#elif defined(MEDIAPIPE_IOS)
   MPPMetalHelper* gpu_helper_ = nullptr;
   std::unique_ptr<GPUData> gpu_data_in_;
   std::vector<std::unique_ptr<GPUData>> gpu_data_out_;
   TFLBufferConvert* converter_from_BPHWC4_ = nil;
 #endif
 
+#if defined(MEDIAPIPE_EDGE_TPU)
+  std::shared_ptr<edgetpu::EdgeTpuContext> edgetpu_context_ =
+      edgetpu::EdgeTpuManager::GetSingleton()->OpenDevice();
+#endif
+
   std::string model_path_ = "";
   bool gpu_inference_ = false;
   bool gpu_input_ = false;
@@ -210,10 +238,9 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
   use_gpu |= options.use_gpu();
 
   if (use_gpu) {
-#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__) && \
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-    !defined(__APPLE__)
     MP_RETURN_IF_ERROR(mediapipe::GlCalculatorHelper::UpdateContract(cc));
-#elif defined(__APPLE__) && !TARGET_OS_OSX  // iOS
+#elif defined(MEDIAPIPE_IOS)
     MP_RETURN_IF_ERROR([MPPMetalHelper updateContract:cc]);
 #endif
   }
@@ -253,26 +280,24 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
   MP_RETURN_IF_ERROR(LoadModel(cc));
 
   if (gpu_inference_) {
-#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__) && \
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-    !defined(__APPLE__)
     MP_RETURN_IF_ERROR(gpu_helper_.Open(cc));
-#elif defined(__APPLE__) && !TARGET_OS_OSX  // iOS
+#elif defined(MEDIAPIPE_IOS)
     gpu_helper_ = [[MPPMetalHelper alloc] initWithCalculatorContext:cc];
     RET_CHECK(gpu_helper_);
 #endif
-#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__) && \
+
-    !defined(__APPLE__)
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
     MP_RETURN_IF_ERROR(gpu_helper_.RunInGlContext(
         [this, &cc]() -> ::mediapipe::Status { return LoadDelegate(cc); }));
 #else
     MP_RETURN_IF_ERROR(LoadDelegate(cc));
 #endif
+  } else {
+#if defined(__EMSCRIPTEN__) || defined(MEDIAPIPE_ANDROID)
+    MP_RETURN_IF_ERROR(LoadDelegate(cc));
+#endif  // __EMSCRIPTEN__ || ANDROID
   }
-
-#if defined(__EMSCRIPTEN__)
-  MP_RETURN_IF_ERROR(LoadDelegate(cc));
-#endif  // __EMSCRIPTEN__
-
   return ::mediapipe::OkStatus();
 }
 
@@ -280,8 +305,7 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
   // 1. Receive pre-processed tensor inputs.
   if (gpu_input_) {
     // Read GPU input into SSBO.
-#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__) && \
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-    !defined(__APPLE__)
     const auto& input_tensors =
         cc->Inputs().Tag("TENSORS_GPU").Get<std::vector<GpuTensor>>();
     RET_CHECK_EQ(input_tensors.size(), 1);
@@ -291,7 +315,7 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
           RET_CHECK_CALL(CopyBuffer(input_tensors[0], gpu_data_in_->buffer));
           return ::mediapipe::OkStatus();
         }));
-#elif defined(__APPLE__) && !TARGET_OS_OSX  // iOS
+#elif defined(MEDIAPIPE_IOS)
     const auto& input_tensors =
         cc->Inputs().Tag("TENSORS_GPU").Get<std::vector<GpuTensor>>();
     RET_CHECK_EQ(input_tensors.size(), 1);
@@ -327,14 +351,13 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
 
   // 2. Run inference.
   if (gpu_inference_) {
-#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__) && \
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-    !defined(__APPLE__)
     MP_RETURN_IF_ERROR(
         gpu_helper_.RunInGlContext([this]() -> ::mediapipe::Status {
           RET_CHECK_EQ(interpreter_->Invoke(), kTfLiteOk);
           return ::mediapipe::OkStatus();
         }));
-#elif defined(__APPLE__) && !TARGET_OS_OSX  // iOS
+#elif defined(MEDIAPIPE_IOS)
     RET_CHECK_EQ(interpreter_->Invoke(), kTfLiteOk);
 #endif
   } else {
@@ -343,8 +366,7 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
 
   // 3. Output processed tensors.
   if (gpu_output_) {
-#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__) && \
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-    !defined(__APPLE__)
     // Output result tensors (GPU).
     auto output_tensors = absl::make_unique<std::vector<GpuTensor>>();
     MP_RETURN_IF_ERROR(gpu_helper_.RunInGlContext(
@@ -361,7 +383,7 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
     cc->Outputs()
         .Tag("TENSORS_GPU")
         .Add(output_tensors.release(), cc->InputTimestamp());
-#elif defined(__APPLE__) && !TARGET_OS_OSX  // iOS
+#elif defined(MEDIAPIPE_IOS)
     // Output result tensors (GPU).
     auto output_tensors = absl::make_unique<std::vector<GpuTensor>>();
     output_tensors->resize(gpu_data_out_.size());
@@ -406,25 +428,29 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
 
 ::mediapipe::Status TfLiteInferenceCalculator::Close(CalculatorContext* cc) {
   if (delegate_) {
-#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__) && \
+    if (gpu_inference_) {
-    !defined(__APPLE__)
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-    MP_RETURN_IF_ERROR(gpu_helper_.RunInGlContext([this]() -> Status {
+      MP_RETURN_IF_ERROR(gpu_helper_.RunInGlContext([this]() -> Status {
-      TfLiteGpuDelegateDelete(delegate_);
+        TfLiteGpuDelegateDelete(delegate_);
+        gpu_data_in_.reset();
+        for (int i = 0; i < gpu_data_out_.size(); ++i) {
+          gpu_data_out_[i].reset();
+        }
+        return ::mediapipe::OkStatus();
+      }));
+#elif defined(MEDIAPIPE_IOS)
+      TFLGpuDelegateDelete(delegate_);
       gpu_data_in_.reset();
       for (int i = 0; i < gpu_data_out_.size(); ++i) {
         gpu_data_out_[i].reset();
       }
-      return ::mediapipe::OkStatus();
-    }));
-#elif defined(__APPLE__) && !TARGET_OS_OSX  // iOS
-    TFLGpuDelegateDelete(delegate_);
-    gpu_data_in_.reset();
-    for (int i = 0; i < gpu_data_out_.size(); ++i) {
-      gpu_data_out_[i].reset();
-    }
 #endif
+    }
     delegate_ = nullptr;
   }
+#if defined(MEDIAPIPE_EDGE_TPU)
+  edgetpu_context_.reset();
+#endif
   return ::mediapipe::OkStatus();
 }
 
@@ -458,16 +484,18 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
   model_ = tflite::FlatBufferModel::BuildFromFile(model_path_.c_str());
   RET_CHECK(model_);
 
+  tflite::ops::builtin::BuiltinOpResolver op_resolver;
   if (cc->InputSidePackets().HasTag("CUSTOM_OP_RESOLVER")) {
-    const auto& op_resolver =
+    op_resolver = cc->InputSidePackets()
-        cc->InputSidePackets()
+                      .Tag("CUSTOM_OP_RESOLVER")
-            .Tag("CUSTOM_OP_RESOLVER")
+                      .Get<tflite::ops::builtin::BuiltinOpResolver>();
-            .Get<tflite::ops::builtin::BuiltinOpResolver>();
-    tflite::InterpreterBuilder(*model_, op_resolver)(&interpreter_);
-  } else {
-    const tflite::ops::builtin::BuiltinOpResolver op_resolver;
-    tflite::InterpreterBuilder(*model_, op_resolver)(&interpreter_);
   }
+#if defined(MEDIAPIPE_EDGE_TPU)
+  interpreter_ =
+      BuildEdgeTpuInterpreter(*model_, &op_resolver, edgetpu_context_.get());
+#else
+  tflite::InterpreterBuilder(*model_, op_resolver)(&interpreter_);
+#endif  // MEDIAPIPE_EDGE_TPU
 
   RET_CHECK(interpreter_);
 
@@ -490,8 +518,22 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
 
 ::mediapipe::Status TfLiteInferenceCalculator::LoadDelegate(
     CalculatorContext* cc) {
-#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__) && \
+#if defined(MEDIAPIPE_ANDROID)
-    !defined(__APPLE__)
+  if (!gpu_inference_) {
+    if (cc->Options<mediapipe::TfLiteInferenceCalculatorOptions>()
+            .use_nnapi()) {
+      // 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);
+    }
+    // Return, no need for GPU delegate below.
+    return ::mediapipe::OkStatus();
+  }
+#endif  // ANDROID
+
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
   // Configure and create the delegate.
   TfLiteGpuDelegateOptions options = TfLiteGpuDelegateOptionsDefault();
   options.compile_options.precision_loss_allowed = 1;
@@ -551,11 +593,11 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
   RET_CHECK_EQ(interpreter_->ModifyGraphWithDelegate(delegate_), kTfLiteOk);
 #endif  // OpenGL
 
-#if defined(__APPLE__) && !TARGET_OS_OSX  // iOS
+#if defined(MEDIAPIPE_IOS)
   // Configure and create the delegate.
-  GpuDelegateOptions options;
+  TFLGpuDelegateOptions options;
   options.allow_precision_loss = false;  // Must match converter, F=float/T=half
-  options.wait_type = GpuDelegateOptions::WaitType::kPassive;
+  options.wait_type = TFLGpuDelegateWaitType::TFLGpuDelegateWaitTypePassive;
   if (!delegate_) delegate_ = TFLGpuDelegateCreate(&options);
   id<MTLDevice> device = gpu_helper_.mtlDevice;
 

mediapipe/calculators/tflite/tflite_inference_calculator.proto

@@ -45,4 +45,9 @@ message TfLiteInferenceCalculatorOptions {
   // input tensors are on CPU. For input tensors on GPU, GPU backend is always
   // used.
   optional bool use_gpu = 2 [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];
 }

mediapipe/calculators/tflite/tflite_tensors_to_classification_calculator.cc

@@ -24,8 +24,7 @@
 #include "mediapipe/framework/port/ret_check.h"
 #include "mediapipe/util/resource_util.h"
 #include "tensorflow/lite/interpreter.h"
-#if defined(__EMSCRIPTEN__) || defined(__ANDROID__) || \
+#if defined(MEDIAPIPE_MOBILE)
-    (defined(__APPLE__) && !TARGET_OS_OSX)
 #include "mediapipe/util/android/file/base/file.h"
 #include "mediapipe/util/android/file/base/helpers.h"
 #else

mediapipe/calculators/tflite/tflite_tensors_to_detections_calculator.cc

@@ -27,8 +27,7 @@
 #include "mediapipe/framework/port/ret_check.h"
 #include "tensorflow/lite/interpreter.h"
 
-#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__) && \
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-    !defined(__APPLE__)
 #include "mediapipe/gpu/gl_calculator_helper.h"
 #include "tensorflow/lite/delegates/gpu/gl/gl_buffer.h"
 #include "tensorflow/lite/delegates/gpu/gl/gl_program.h"
@@ -36,7 +35,7 @@
 #include "tensorflow/lite/delegates/gpu/gl_delegate.h"
 #endif  // !MEDIAPIPE_DISABLE_GPU
 
-#if defined(__APPLE__) && !TARGET_OS_OSX  // iOS
+#if defined(MEDIAPIPE_IOS)
 #import <CoreVideo/CoreVideo.h>
 #import <Metal/Metal.h>
 #import <MetalKit/MetalKit.h>
@@ -56,17 +55,15 @@ constexpr int kNumCoordsPerBox = 4;
 
 namespace mediapipe {
 
-#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__) && \
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-    !defined(__APPLE__)
 using ::tflite::gpu::gl::CreateReadWriteShaderStorageBuffer;
 using ::tflite::gpu::gl::GlShader;
 #endif
 
-#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__) && \
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-    !defined(__APPLE__)
 typedef ::tflite::gpu::gl::GlBuffer GpuTensor;
 typedef ::tflite::gpu::gl::GlProgram GpuProgram;
-#elif defined(__APPLE__) && !TARGET_OS_OSX  // iOS
+#elif defined(MEDIAPIPE_IOS)
 typedef id<MTLBuffer> GpuTensor;
 typedef id<MTLComputePipelineState> GpuProgram;
 #endif
@@ -172,24 +169,21 @@ class TfLiteTensorsToDetectionsCalculator : public CalculatorBase {
       const int* detection_classes, std::vector<Detection>* output_detections);
   Detection ConvertToDetection(float box_ymin, float box_xmin, float box_ymax,
                                float box_xmax, float score, int class_id,
-                               int detection_id, bool flip_vertically);
+                               bool flip_vertically);
 
   int num_classes_ = 0;
   int num_boxes_ = 0;
   int num_coords_ = 0;
-  // Unique detection ID per new detection.
-  static int next_detection_id_;
   std::set<int> ignore_classes_;
 
   ::mediapipe::TfLiteTensorsToDetectionsCalculatorOptions options_;
   std::vector<Anchor> anchors_;
   bool side_packet_anchors_{};
 
-#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__) && \
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-    !defined(__APPLE__)
   mediapipe::GlCalculatorHelper gpu_helper_;
   std::unique_ptr<GPUData> gpu_data_;
-#elif defined(__APPLE__) && !TARGET_OS_OSX  // iOS
+#elif defined(MEDIAPIPE_IOS)
   MPPMetalHelper* gpu_helper_ = nullptr;
   std::unique_ptr<GPUData> gpu_data_;
 #endif
@@ -199,10 +193,6 @@ class TfLiteTensorsToDetectionsCalculator : public CalculatorBase {
 };
 REGISTER_CALCULATOR(TfLiteTensorsToDetectionsCalculator);
 
-// Initialization of non-const static member should happen outside class
-// definition.
-int TfLiteTensorsToDetectionsCalculator::next_detection_id_ = 0;
-
 ::mediapipe::Status TfLiteTensorsToDetectionsCalculator::GetContract(
     CalculatorContract* cc) {
   RET_CHECK(!cc->Inputs().GetTags().empty());
@@ -232,10 +222,9 @@ int TfLiteTensorsToDetectionsCalculator::next_detection_id_ = 0;
   }
 
   if (use_gpu) {
-#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__) && \
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-    !defined(__APPLE__)
     MP_RETURN_IF_ERROR(mediapipe::GlCalculatorHelper::UpdateContract(cc));
-#elif defined(__APPLE__) && !TARGET_OS_OSX  // iOS
+#elif defined(MEDIAPIPE_IOS)
     MP_RETURN_IF_ERROR([MPPMetalHelper updateContract:cc]);
 #endif
   }
@@ -249,10 +238,9 @@ int TfLiteTensorsToDetectionsCalculator::next_detection_id_ = 0;
 
   if (cc->Inputs().HasTag("TENSORS_GPU")) {
     gpu_input_ = true;
-#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__) && \
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-    !defined(__APPLE__)
     MP_RETURN_IF_ERROR(gpu_helper_.Open(cc));
-#elif defined(__APPLE__) && !TARGET_OS_OSX  // iOS
+#elif defined(MEDIAPIPE_IOS)
     gpu_helper_ = [[MPPMetalHelper alloc] initWithCalculatorContext:cc];
     RET_CHECK(gpu_helper_);
 #endif
@@ -412,8 +400,7 @@ int TfLiteTensorsToDetectionsCalculator::next_detection_id_ = 0;
 }
 ::mediapipe::Status TfLiteTensorsToDetectionsCalculator::ProcessGPU(
     CalculatorContext* cc, std::vector<Detection>* output_detections) {
-#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__) && \
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-    !defined(__APPLE__)
   const auto& input_tensors =
       cc->Inputs().Tag("TENSORS_GPU").Get<std::vector<GpuTensor>>();
   RET_CHECK_GE(input_tensors.size(), 2);
@@ -476,7 +463,7 @@ int TfLiteTensorsToDetectionsCalculator::next_detection_id_ = 0;
 
     return ::mediapipe::OkStatus();
   }));
-#elif defined(__APPLE__) && !TARGET_OS_OSX  // iOS
+#elif defined(MEDIAPIPE_IOS)
 
   const auto& input_tensors =
       cc->Inputs().Tag("TENSORS_GPU").Get<std::vector<GpuTensor>>();
@@ -575,12 +562,11 @@ int TfLiteTensorsToDetectionsCalculator::next_detection_id_ = 0;
 
 ::mediapipe::Status TfLiteTensorsToDetectionsCalculator::Close(
     CalculatorContext* cc) {
-#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__) && \
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-    !defined(__APPLE__)
   gpu_helper_.RunInGlContext([this] { gpu_data_.reset(); });
-#elif defined(__APPLE__) && !TARGET_OS_OSX  // iOS
+#elif defined(MEDIAPIPE_IOS)
   gpu_data_.reset();
-#endif                                      //  !MEDIAPIPE_DISABLE_GPU
+#endif
 
   return ::mediapipe::OkStatus();
 }
@@ -686,10 +672,7 @@ int TfLiteTensorsToDetectionsCalculator::next_detection_id_ = 0;
     Detection detection = ConvertToDetection(
         detection_boxes[box_offset + 0], detection_boxes[box_offset + 1],
         detection_boxes[box_offset + 2], detection_boxes[box_offset + 3],
-        detection_scores[i], detection_classes[i], next_detection_id_,
+        detection_scores[i], detection_classes[i], options_.flip_vertically());
-        options_.flip_vertically());
-    // Increment to get next unique detection ID.
-    ++next_detection_id_;
     // Add keypoints.
     if (options_.num_keypoints() > 0) {
       auto* location_data = detection.mutable_location_data();
@@ -712,11 +695,10 @@ int TfLiteTensorsToDetectionsCalculator::next_detection_id_ = 0;
 
 Detection TfLiteTensorsToDetectionsCalculator::ConvertToDetection(
     float box_ymin, float box_xmin, float box_ymax, float box_xmax, float score,
-    int class_id, int detection_id, bool flip_vertically) {
+    int class_id, bool flip_vertically) {
   Detection detection;
   detection.add_score(score);
   detection.add_label_id(class_id);
-  detection.set_detection_id(detection_id);
 
   LocationData* location_data = detection.mutable_location_data();
   location_data->set_format(LocationData::RELATIVE_BOUNDING_BOX);
@@ -733,8 +715,7 @@ Detection TfLiteTensorsToDetectionsCalculator::ConvertToDetection(
 
 ::mediapipe::Status TfLiteTensorsToDetectionsCalculator::GpuInit(
     CalculatorContext* cc) {
-#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__) && \
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-    !defined(__APPLE__)
   MP_RETURN_IF_ERROR(gpu_helper_.RunInGlContext([this]()
                                                     -> ::mediapipe::Status {
     gpu_data_ = absl::make_unique<GPUData>();
@@ -947,8 +928,7 @@ void main() {
     return ::mediapipe::OkStatus();
   }));
 
-#elif defined(__APPLE__) && !TARGET_OS_OSX  // iOS
+#elif defined(MEDIAPIPE_IOS)
-  // TODO consolidate Metal and OpenGL shaders via vulkan.
 
   gpu_data_ = absl::make_unique<GPUData>();
   id<MTLDevice> device = gpu_helper_.mtlDevice;
@@ -1178,7 +1158,7 @@ kernel void scoreKernel(
     CHECK_LT(num_classes_, max_wg_size) << "# classes must be <" << max_wg_size;
   }
 
-#endif  // __ANDROID__ or iOS
+#endif  // !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
 
   return ::mediapipe::OkStatus();
 }

mediapipe/calculators/tflite/tflite_tensors_to_landmarks_calculator.cc

@@ -76,11 +76,11 @@ REGISTER_CALCULATOR(TfLiteTensorsToLandmarksCalculator);
   }
 
   if (cc->Outputs().HasTag("LANDMARKS")) {
-    cc->Outputs().Tag("LANDMARKS").Set<std::vector<Landmark>>();
+    cc->Outputs().Tag("LANDMARKS").Set<LandmarkList>();
   }
 
   if (cc->Outputs().HasTag("NORM_LANDMARKS")) {
-    cc->Outputs().Tag("NORM_LANDMARKS").Set<std::vector<NormalizedLandmark>>();
+    cc->Outputs().Tag("NORM_LANDMARKS").Set<NormalizedLandmarkList>();
   }
 
   return ::mediapipe::OkStatus();
@@ -127,54 +127,55 @@ REGISTER_CALCULATOR(TfLiteTensorsToLandmarksCalculator);
 
   const float* raw_landmarks = raw_tensor->data.f;
 
-  auto output_landmarks = absl::make_unique<std::vector<Landmark>>();
+  LandmarkList output_landmarks;
 
   for (int ld = 0; ld < num_landmarks_; ++ld) {
     const int offset = ld * num_dimensions;
-    Landmark landmark;
+    Landmark* landmark = output_landmarks.add_landmark();
 
     if (options_.flip_horizontally()) {
-      landmark.set_x(options_.input_image_width() - raw_landmarks[offset]);
+      landmark->set_x(options_.input_image_width() - raw_landmarks[offset]);
     } else {
-      landmark.set_x(raw_landmarks[offset]);
+      landmark->set_x(raw_landmarks[offset]);
     }
     if (num_dimensions > 1) {
       if (options_.flip_vertically()) {
-        landmark.set_y(options_.input_image_height() -
+        landmark->set_y(options_.input_image_height() -
-                       raw_landmarks[offset + 1]);
+                        raw_landmarks[offset + 1]);
       } else {
-        landmark.set_y(raw_landmarks[offset + 1]);
+        landmark->set_y(raw_landmarks[offset + 1]);
       }
     }
     if (num_dimensions > 2) {
-      landmark.set_z(raw_landmarks[offset + 2]);
+      landmark->set_z(raw_landmarks[offset + 2]);
     }
-    output_landmarks->push_back(landmark);
   }
 
   // Output normalized landmarks if required.
   if (cc->Outputs().HasTag("NORM_LANDMARKS")) {
-    auto output_norm_landmarks =
+    NormalizedLandmarkList output_norm_landmarks;
-        absl::make_unique<std::vector<NormalizedLandmark>>();
+    // for (const auto& landmark : output_landmarks) {
-    for (const auto& landmark : *output_landmarks) {
+    for (int i = 0; i < output_landmarks.landmark_size(); ++i) {
-      NormalizedLandmark norm_landmark;
+      const Landmark& landmark = output_landmarks.landmark(i);
-      norm_landmark.set_x(static_cast<float>(landmark.x()) /
+      NormalizedLandmark* norm_landmark = output_norm_landmarks.add_landmark();
-                          options_.input_image_width());
+      norm_landmark->set_x(static_cast<float>(landmark.x()) /
-      norm_landmark.set_y(static_cast<float>(landmark.y()) /
+                           options_.input_image_width());
-                          options_.input_image_height());
+      norm_landmark->set_y(static_cast<float>(landmark.y()) /
-      norm_landmark.set_z(landmark.z() / options_.normalize_z());
+                           options_.input_image_height());
-
+      norm_landmark->set_z(landmark.z() / options_.normalize_z());
-      output_norm_landmarks->push_back(norm_landmark);
     }
     cc->Outputs()
         .Tag("NORM_LANDMARKS")
-        .Add(output_norm_landmarks.release(), cc->InputTimestamp());
+        .AddPacket(MakePacket<NormalizedLandmarkList>(output_norm_landmarks)
+                       .At(cc->InputTimestamp()));
   }
+
   // Output absolute landmarks.
   if (cc->Outputs().HasTag("LANDMARKS")) {
     cc->Outputs()
         .Tag("LANDMARKS")
-        .Add(output_landmarks.release(), cc->InputTimestamp());
+        .AddPacket(MakePacket<LandmarkList>(output_landmarks)
+                       .At(cc->InputTimestamp()));
   }
 
   return ::mediapipe::OkStatus();

mediapipe/calculators/tflite/tflite_tensors_to_segmentation_calculator.cc

@@ -28,8 +28,7 @@
 #include "mediapipe/util/resource_util.h"
 #include "tensorflow/lite/interpreter.h"
 
-#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__) && \
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-    !defined(__APPLE__)
 #include "mediapipe/gpu/gl_calculator_helper.h"
 #include "mediapipe/gpu/gl_simple_shaders.h"
 #include "mediapipe/gpu/shader_util.h"
@@ -54,8 +53,7 @@ float Clamp(float val, float min, float max) {
 
 namespace mediapipe {
 
-#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__) && \
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-    !defined(__APPLE__)
 using ::tflite::gpu::gl::CopyBuffer;
 using ::tflite::gpu::gl::CreateReadWriteRgbaImageTexture;
 using ::tflite::gpu::gl::CreateReadWriteShaderStorageBuffer;
@@ -131,8 +129,7 @@ class TfLiteTensorsToSegmentationCalculator : public CalculatorBase {
   int tensor_channels_ = 0;
 
   bool use_gpu_ = false;
-#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__) && \
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-    !defined(__APPLE__)
   mediapipe::GlCalculatorHelper gpu_helper_;
   std::unique_ptr<GlProgram> mask_program_with_prev_;
   std::unique_ptr<GlProgram> mask_program_no_prev_;
@@ -162,8 +159,7 @@ REGISTER_CALCULATOR(TfLiteTensorsToSegmentationCalculator);
   }
 
   // Inputs GPU.
-#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__) && \
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-    !defined(__APPLE__)
   if (cc->Inputs().HasTag("TENSORS_GPU")) {
     cc->Inputs().Tag("TENSORS_GPU").Set<std::vector<GlBuffer>>();
     use_gpu |= true;
@@ -182,8 +178,7 @@ REGISTER_CALCULATOR(TfLiteTensorsToSegmentationCalculator);
   if (cc->Outputs().HasTag("MASK")) {
     cc->Outputs().Tag("MASK").Set<ImageFrame>();
   }
-#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__) && \
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-    !defined(__APPLE__)
   if (cc->Outputs().HasTag("MASK_GPU")) {
     cc->Outputs().Tag("MASK_GPU").Set<mediapipe::GpuBuffer>();
     use_gpu |= true;
@@ -191,8 +186,7 @@ REGISTER_CALCULATOR(TfLiteTensorsToSegmentationCalculator);
 #endif  //  !MEDIAPIPE_DISABLE_GPU
 
   if (use_gpu) {
-#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__) && \
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-    !defined(__APPLE__)
     MP_RETURN_IF_ERROR(mediapipe::GlCalculatorHelper::UpdateContract(cc));
 #endif  //  !MEDIAPIPE_DISABLE_GPU
   }
@@ -205,8 +199,7 @@ REGISTER_CALCULATOR(TfLiteTensorsToSegmentationCalculator);
 
   if (cc->Inputs().HasTag("TENSORS_GPU")) {
     use_gpu_ = true;
-#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__) && \
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-    !defined(__APPLE__)
     MP_RETURN_IF_ERROR(gpu_helper_.Open(cc));
 #endif  //  !MEDIAPIPE_DISABLE_GPU
   }
@@ -214,8 +207,7 @@ REGISTER_CALCULATOR(TfLiteTensorsToSegmentationCalculator);
   MP_RETURN_IF_ERROR(LoadOptions(cc));
 
   if (use_gpu_) {
-#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__) && \
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-    !defined(__APPLE__)
     MP_RETURN_IF_ERROR(
         gpu_helper_.RunInGlContext([this, cc]() -> ::mediapipe::Status {
           MP_RETURN_IF_ERROR(InitGpu(cc));
@@ -232,8 +224,7 @@ REGISTER_CALCULATOR(TfLiteTensorsToSegmentationCalculator);
 ::mediapipe::Status TfLiteTensorsToSegmentationCalculator::Process(
     CalculatorContext* cc) {
   if (use_gpu_) {
-#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__) && \
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-    !defined(__APPLE__)
     MP_RETURN_IF_ERROR(
         gpu_helper_.RunInGlContext([this, cc]() -> ::mediapipe::Status {
           MP_RETURN_IF_ERROR(ProcessGpu(cc));
@@ -249,8 +240,7 @@ REGISTER_CALCULATOR(TfLiteTensorsToSegmentationCalculator);
 
 ::mediapipe::Status TfLiteTensorsToSegmentationCalculator::Close(
     CalculatorContext* cc) {
-#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__) && \
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-    !defined(__APPLE__)
   gpu_helper_.RunInGlContext([this] {
     if (upsample_program_) glDeleteProgram(upsample_program_);
     upsample_program_ = 0;
@@ -377,8 +367,7 @@ REGISTER_CALCULATOR(TfLiteTensorsToSegmentationCalculator);
   if (cc->Inputs().Tag("TENSORS_GPU").IsEmpty()) {
     return ::mediapipe::OkStatus();
   }
-#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__) && \
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-    !defined(__APPLE__)
   // Get input streams.
   const auto& input_tensors =
       cc->Inputs().Tag("TENSORS_GPU").Get<std::vector<GlBuffer>>();
@@ -464,8 +453,7 @@ REGISTER_CALCULATOR(TfLiteTensorsToSegmentationCalculator);
 }
 
 void TfLiteTensorsToSegmentationCalculator::GlRender() {
-#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__) && \
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-    !defined(__APPLE__)
   static const GLfloat square_vertices[] = {
       -1.0f, -1.0f,  // bottom left
       1.0f,  -1.0f,  // bottom right
@@ -537,8 +525,7 @@ void TfLiteTensorsToSegmentationCalculator::GlRender() {
 
 ::mediapipe::Status TfLiteTensorsToSegmentationCalculator::InitGpu(
     CalculatorContext* cc) {
-#if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__) && \
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-    !defined(__APPLE__)
   MP_RETURN_IF_ERROR(gpu_helper_.RunInGlContext([this]()
                                                     -> ::mediapipe::Status {
     // A shader to process a segmentation tensor into an output mask,

mediapipe/calculators/util/BUILD

@@ -12,14 +12,14 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+load("//mediapipe/framework/port:build_config.bzl", "mediapipe_cc_proto_library")
+
 licenses(["notice"])  # Apache 2.0
 
 package(default_visibility = ["//visibility:public"])
 
 exports_files(["LICENSE"])
 
-load("//mediapipe/framework/port:build_config.bzl", "mediapipe_cc_proto_library")
-
 proto_library(
     name = "annotation_overlay_calculator_proto",
     srcs = ["annotation_overlay_calculator.proto"],
@@ -72,6 +72,24 @@ proto_library(
     ],
 )
 
+proto_library(
+    name = "collection_has_min_size_calculator_proto",
+    srcs = ["collection_has_min_size_calculator.proto"],
+    visibility = ["//visibility:public"],
+    deps = [
+        "//mediapipe/framework:calculator_proto",
+    ],
+)
+
+proto_library(
+    name = "association_calculator_proto",
+    srcs = ["association_calculator.proto"],
+    visibility = ["//visibility:public"],
+    deps = [
+        "//mediapipe/framework:calculator_proto",
+    ],
+)
+
 mediapipe_cc_proto_library(
     name = "annotation_overlay_calculator_cc_proto",
     srcs = ["annotation_overlay_calculator.proto"],
@@ -141,6 +159,26 @@ mediapipe_cc_proto_library(
     ],
 )
 
+mediapipe_cc_proto_library(
+    name = "collection_has_min_size_calculator_cc_proto",
+    srcs = ["collection_has_min_size_calculator.proto"],
+    cc_deps = [
+        "//mediapipe/framework:calculator_cc_proto",
+    ],
+    visibility = ["//mediapipe:__subpackages__"],
+    deps = [":collection_has_min_size_calculator_proto"],
+)
+
+mediapipe_cc_proto_library(
+    name = "association_calculator_cc_proto",
+    srcs = ["association_calculator.proto"],
+    cc_deps = [
+        "//mediapipe/framework:calculator_cc_proto",
+    ],
+    visibility = ["//mediapipe:__subpackages__"],
+    deps = [":association_calculator_proto"],
+)
+
 cc_library(
     name = "packet_frequency_calculator",
     srcs = ["packet_frequency_calculator.cc"],
@@ -847,3 +885,101 @@ cc_library(
     ],
     alwayslink = 1,
 )
+
+cc_library(
+    name = "filter_collection_calculator",
+    srcs = ["filter_collection_calculator.cc"],
+    hdrs = ["filter_collection_calculator.h"],
+    visibility = ["//visibility:public"],
+    deps = [
+        "//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",
+        "@com_google_absl//absl/strings",
+    ],
+    alwayslink = 1,
+)
+
+cc_library(
+    name = "collection_has_min_size_calculator",
+    srcs = ["collection_has_min_size_calculator.cc"],
+    hdrs = ["collection_has_min_size_calculator.h"],
+    visibility = ["//visibility:public"],
+    deps = [
+        ":collection_has_min_size_calculator_cc_proto",
+        "//mediapipe/framework:calculator_framework",
+        "//mediapipe/framework/formats:rect_cc_proto",
+        "//mediapipe/framework/port:ret_check",
+        "//mediapipe/framework/port:status",
+    ],
+    alwayslink = 1,
+)
+
+cc_library(
+    name = "association_calculator",
+    hdrs = ["association_calculator.h"],
+    visibility = ["//visibility:public"],
+    deps = [
+        ":association_calculator_cc_proto",
+        "//mediapipe/framework:calculator_context",
+        "//mediapipe/framework:calculator_framework",
+        "//mediapipe/framework:collection_item_id",
+        "//mediapipe/framework/port:rectangle",
+        "//mediapipe/framework/port:status",
+        "@com_google_absl//absl/memory",
+    ],
+    alwayslink = 1,
+)
+
+cc_library(
+    name = "association_norm_rect_calculator",
+    srcs = ["association_norm_rect_calculator.cc"],
+    visibility = ["//visibility:public"],
+    deps = [
+        ":association_calculator",
+        "//mediapipe/framework:calculator_context",
+        "//mediapipe/framework:calculator_framework",
+        "//mediapipe/framework/formats:rect_cc_proto",
+        "//mediapipe/framework/port:rectangle",
+        "//mediapipe/framework/port:status",
+    ],
+    alwayslink = 1,
+)
+
+cc_library(
+    name = "association_detection_calculator",
+    srcs = ["association_detection_calculator.cc"],
+    visibility = ["//visibility:public"],
+    deps = [
+        ":association_calculator",
+        "//mediapipe/framework:calculator_context",
+        "//mediapipe/framework:calculator_framework",
+        "//mediapipe/framework/formats:detection_cc_proto",
+        "//mediapipe/framework/formats:location",
+        "//mediapipe/framework/port:rectangle",
+        "//mediapipe/framework/port:status",
+    ],
+    alwayslink = 1,
+)
+
+cc_test(
+    name = "association_calculator_test",
+    srcs = ["association_calculator_test.cc"],
+    deps = [
+        ":association_detection_calculator",
+        ":association_norm_rect_calculator",
+        "//mediapipe/framework:calculator_cc_proto",
+        "//mediapipe/framework:calculator_framework",
+        "//mediapipe/framework:calculator_runner",
+        "//mediapipe/framework:collection_item_id",
+        "//mediapipe/framework:packet",
+        "//mediapipe/framework/deps:message_matchers",
+        "//mediapipe/framework/formats:detection_cc_proto",
+        "//mediapipe/framework/formats:location_data_cc_proto",
+        "//mediapipe/framework/formats:rect_cc_proto",
+        "//mediapipe/framework/port:gtest_main",
+        "//mediapipe/framework/port:parse_text_proto",
+    ],
+)

mediapipe/calculators/util/annotation_overlay_calculator.cc

@@ -148,9 +148,6 @@ class AnnotationOverlayCalculator : public CalculatorBase {
   // Underlying helper renderer library.
   std::unique_ptr<AnnotationRenderer> renderer_;
 
-  // Number of input streams with render data.
-  int num_render_streams_;
-
   // Indicates if image frame is available as input.
   bool image_frame_available_ = false;
 
@@ -181,20 +178,15 @@ REGISTER_CALCULATOR(AnnotationOverlayCalculator);
     return ::mediapipe::InternalError("GPU output must have GPU input.");
   }
 
-  // Assume all inputs are render streams; adjust below.
-  int num_render_streams = cc->Inputs().NumEntries();
-
   // Input image to render onto copy of.
 #if !defined(MEDIAPIPE_DISABLE_GPU)
   if (cc->Inputs().HasTag(kInputFrameTagGpu)) {
     cc->Inputs().Tag(kInputFrameTagGpu).Set<mediapipe::GpuBuffer>();
-    num_render_streams = cc->Inputs().NumEntries() - 1;
     use_gpu |= true;
   }
 #endif  //  !MEDIAPIPE_DISABLE_GPU
   if (cc->Inputs().HasTag(kInputFrameTag)) {
     cc->Inputs().Tag(kInputFrameTag).Set<ImageFrame>();
-    num_render_streams = cc->Inputs().NumEntries() - 1;
   }
 
   // Data streams to render.
@@ -246,12 +238,10 @@ REGISTER_CALCULATOR(AnnotationOverlayCalculator);
   if (cc->Inputs().HasTag(kInputFrameTagGpu) ||
       cc->Inputs().HasTag(kInputFrameTag)) {
     image_frame_available_ = true;
-    num_render_streams_ = cc->Inputs().NumEntries() - 1;
   } else {
     image_frame_available_ = false;
     RET_CHECK(options_.has_canvas_width_px());
     RET_CHECK(options_.has_canvas_height_px());
-    num_render_streams_ = cc->Inputs().NumEntries();
   }
 
   // Initialize the helper renderer library.

mediapipe/calculators/util/association_calculator.h

@@ -0,0 +1,259 @@
+// 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.
+
+#ifndef MEDIAPIPE_CALCULATORS_UTIL_ASSOCIATION_CALCULATOR_H_
+#define MEDIAPIPE_CALCULATORS_UTIL_ASSOCIATION_CALCULATOR_H_
+
+#include <memory>
+#include <vector>
+
+#include "absl/memory/memory.h"
+#include "mediapipe/calculators/util/association_calculator.pb.h"
+#include "mediapipe/framework/calculator_context.h"
+#include "mediapipe/framework/calculator_framework.h"
+#include "mediapipe/framework/collection_item_id.h"
+#include "mediapipe/framework/port/canonical_errors.h"
+#include "mediapipe/framework/port/rectangle.h"
+#include "mediapipe/framework/port/status.h"
+
+namespace mediapipe {
+
+// Computes the overlap similarity based on Intersection over Union (IoU) of
+// two rectangles.
+inline float OverlapSimilarity(const Rectangle_f& rect1,
+                               const Rectangle_f& rect2) {
+  if (!rect1.Intersects(rect2)) return 0.0f;
+  // Compute IoU similarity score.
+  const float intersection_area = Rectangle_f(rect1).Intersect(rect2).Area();
+  const float normalization = rect1.Area() + rect2.Area() - intersection_area;
+  return normalization > 0.0f ? intersection_area / normalization : 0.0f;
+}
+
+// AssocationCalculator<T> accepts multiple inputs of vectors of type T that can
+// be converted to Rectangle_f. The output is a vector of type T that contains
+// elements from the input vectors that don't overlap with each other. When
+// two elements overlap, the element that comes in from a later input stream
+// is kept in the output. This association operation is useful for multiple
+// instance inference pipelines in MediaPipe.
+// If an input stream is tagged with "PREV" tag, IDs of overlapping elements
+// from "PREV" input stream are propagated to the output. Elements in the "PREV"
+// input stream that don't overlap with other elements are not added to the
+// output. This stream is designed to take detections from previous timestamp,
+// e.g. output of PreviousLoopbackCalculator to provide temporal association.
+// See AssociationDetectionCalculator and AssociationNormRectCalculator for
+// example uses.
+template <typename T>
+class AssociationCalculator : public CalculatorBase {
+ public:
+  static ::mediapipe::Status GetContract(CalculatorContract* cc) {
+    // Atmost one input stream can be tagged with "PREV".
+    RET_CHECK_LE(cc->Inputs().NumEntries("PREV"), 1);
+
+    if (cc->Inputs().HasTag("PREV")) {
+      RET_CHECK_GE(cc->Inputs().NumEntries(), 2);
+    }
+
+    for (CollectionItemId id = cc->Inputs().BeginId();
+         id < cc->Inputs().EndId(); ++id) {
+      cc->Inputs().Get(id).Set<std::vector<T>>();
+    }
+
+    cc->Outputs().Index(0).Set<std::vector<T>>();
+
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Open(CalculatorContext* cc) override {
+    cc->SetOffset(TimestampDiff(0));
+
+    has_prev_input_stream_ = cc->Inputs().HasTag("PREV");
+    if (has_prev_input_stream_) {
+      prev_input_stream_id_ = cc->Inputs().GetId("PREV", 0);
+    }
+    options_ = cc->Options<::mediapipe::AssociationCalculatorOptions>();
+    CHECK_GE(options_.min_similarity_threshold(), 0);
+
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Process(CalculatorContext* cc) override {
+    auto get_non_overlapping_elements = GetNonOverlappingElements(cc);
+    if (!get_non_overlapping_elements.ok()) {
+      return get_non_overlapping_elements.status();
+    }
+    std::list<T> result = get_non_overlapping_elements.ValueOrDie();
+
+    if (has_prev_input_stream_ &&
+        !cc->Inputs().Get(prev_input_stream_id_).IsEmpty()) {
+      // Processed all regular input streams. Now compare the result list
+      // elements with those in the PREV input stream, and propagate IDs from
+      // PREV input stream as appropriate.
+      const std::vector<T>& prev_input_vec =
+          cc->Inputs()
+              .Get(prev_input_stream_id_)
+              .template Get<std::vector<T>>();
+
+      MP_RETURN_IF_ERROR(
+          PropagateIdsFromPreviousToCurrent(prev_input_vec, &result));
+    }
+
+    auto output = absl::make_unique<std::vector<T>>();
+    for (auto it = result.begin(); it != result.end(); ++it) {
+      output->push_back(*it);
+    }
+    cc->Outputs().Index(0).Add(output.release(), cc->InputTimestamp());
+
+    return ::mediapipe::OkStatus();
+  }
+
+ protected:
+  ::mediapipe::AssociationCalculatorOptions options_;
+
+  bool has_prev_input_stream_;
+  CollectionItemId prev_input_stream_id_;
+
+  virtual ::mediapipe::StatusOr<Rectangle_f> GetRectangle(const T& input) {
+    return ::mediapipe::OkStatus();
+  }
+
+  virtual std::pair<bool, int> GetId(const T& input) { return {false, -1}; }
+
+  virtual void SetId(T* input, int id) {}
+
+ private:
+  // Get a list of non-overlapping elements from all input streams, with
+  // increasing order of priority based on input stream index.
+  mediapipe::StatusOr<std::list<T>> GetNonOverlappingElements(
+      CalculatorContext* cc) {
+    std::list<T> result;
+
+    // Initialize result with the first non-empty input vector.
+    CollectionItemId non_empty_id = cc->Inputs().BeginId();
+    for (CollectionItemId id = cc->Inputs().BeginId();
+         id < cc->Inputs().EndId(); ++id) {
+      if (id == prev_input_stream_id_ || cc->Inputs().Get(id).IsEmpty()) {
+        continue;
+      }
+      const std::vector<T>& input_vec =
+          cc->Inputs().Get(id).Get<std::vector<T>>();
+      if (!input_vec.empty()) {
+        non_empty_id = id;
+        result.push_back(input_vec[0]);
+        for (int j = 1; j < input_vec.size(); ++j) {
+          MP_RETURN_IF_ERROR(AddElementToList(input_vec[j], &result));
+        }
+        break;
+      }
+    }
+
+    // Compare remaining input vectors with the non-empty result vector,
+    // remove lower-priority overlapping elements from the result vector and
+    // had corresponding higher-priority elements as necessary.
+    for (CollectionItemId id = non_empty_id + 1; id < cc->Inputs().EndId();
+         ++id) {
+      if (id == prev_input_stream_id_ || cc->Inputs().Get(id).IsEmpty()) {
+        continue;
+      }
+      const std::vector<T>& input_vec =
+          cc->Inputs().Get(id).Get<std::vector<T>>();
+
+      for (int vi = 0; vi < input_vec.size(); ++vi) {
+        MP_RETURN_IF_ERROR(AddElementToList(input_vec[vi], &result));
+      }
+    }
+
+    return result;
+  }
+
+  ::mediapipe::Status AddElementToList(T element, std::list<T>* current) {
+    // Compare this element with elements of the input collection. If this
+    // element has high overlap with elements of the collection, remove
+    // those elements from the collection and add this element.
+    ASSIGN_OR_RETURN(auto cur_rect, GetRectangle(element));
+
+    bool change_id = false;
+    int new_elem_id = -1;
+
+    for (auto uit = current->begin(); uit != current->end();) {
+      ASSIGN_OR_RETURN(auto prev_rect, GetRectangle(*uit));
+      if (OverlapSimilarity(cur_rect, prev_rect) >
+          options_.min_similarity_threshold()) {
+        std::pair<bool, int> prev_id = GetId(*uit);
+        // If prev_id.first is false when some element doesn't have an ID,
+        // change_id and new_elem_id will not be updated.
+        if (prev_id.first) {
+          change_id = prev_id.first;
+          new_elem_id = prev_id.second;
+        }
+        uit = current->erase(uit);
+      } else {
+        ++uit;
+      }
+    }
+
+    if (change_id) {
+      SetId(&element, new_elem_id);
+    }
+    current->push_back(element);
+
+    return ::mediapipe::OkStatus();
+  }
+
+  // Compare elements of the current list with elements in from the collection
+  // of elements from the previous input stream, and propagate IDs from the
+  // previous input stream as appropriate.
+  ::mediapipe::Status PropagateIdsFromPreviousToCurrent(
+      const std::vector<T>& prev_input_vec, std::list<T>* current) {
+    for (auto vit = current->begin(); vit != current->end(); ++vit) {
+      auto get_cur_rectangle = GetRectangle(*vit);
+      if (!get_cur_rectangle.ok()) {
+        return get_cur_rectangle.status();
+      }
+      const Rectangle_f& cur_rect = get_cur_rectangle.ValueOrDie();
+
+      bool change_id = false;
+      int id_for_vi = -1;
+
+      for (int ui = 0; ui < prev_input_vec.size(); ++ui) {
+        auto get_prev_rectangle = GetRectangle(prev_input_vec[ui]);
+        if (!get_prev_rectangle.ok()) {
+          return get_prev_rectangle.status();
+        }
+        const Rectangle_f& prev_rect = get_prev_rectangle.ValueOrDie();
+
+        if (OverlapSimilarity(cur_rect, prev_rect) >
+            options_.min_similarity_threshold()) {
+          std::pair<bool, int> prev_id = GetId(prev_input_vec[ui]);
+          // If prev_id.first is false when some element doesn't have an ID,
+          // change_id and id_for_vi will not be updated.
+          if (prev_id.first) {
+            change_id = prev_id.first;
+            id_for_vi = prev_id.second;
+          }
+        }
+      }
+
+      if (change_id) {
+        T element = *vit;
+        SetId(&element, id_for_vi);
+        *vit = element;
+      }
+    }
+    return ::mediapipe::OkStatus();
+  }
+};
+
+}  // namespace mediapipe
+
+#endif  // MEDIAPIPE_CALCULATORS_UTIL_ASSOCIATION_CALCULATOR_H_

mediapipe/calculators/util/association_calculator.proto

@@ -0,0 +1,27 @@
+// 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 AssociationCalculatorOptions {
+  extend CalculatorOptions {
+    optional AssociationCalculatorOptions ext = 275124847;
+  }
+
+  optional float min_similarity_threshold = 1 [default = 1.0];
+}

mediapipe/calculators/util/association_calculator_test.cc

@@ -0,0 +1,476 @@
+// 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/framework/calculator.pb.h"
+#include "mediapipe/framework/calculator_framework.h"
+#include "mediapipe/framework/calculator_runner.h"
+#include "mediapipe/framework/collection_item_id.h"
+#include "mediapipe/framework/deps/message_matchers.h"
+#include "mediapipe/framework/formats/detection.pb.h"
+#include "mediapipe/framework/formats/location_data.pb.h"
+#include "mediapipe/framework/formats/rect.pb.h"
+#include "mediapipe/framework/packet.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"
+
+namespace mediapipe {
+
+namespace {
+
+::mediapipe::Detection DetectionWithRelativeLocationData(double xmin,
+                                                         double ymin,
+                                                         double width,
+                                                         double height) {
+  ::mediapipe::Detection detection;
+  ::mediapipe::LocationData* location_data = detection.mutable_location_data();
+  location_data->set_format(::mediapipe::LocationData::RELATIVE_BOUNDING_BOX);
+  location_data->mutable_relative_bounding_box()->set_xmin(xmin);
+  location_data->mutable_relative_bounding_box()->set_ymin(ymin);
+  location_data->mutable_relative_bounding_box()->set_width(width);
+  location_data->mutable_relative_bounding_box()->set_height(height);
+  return detection;
+}
+
+}  // namespace
+
+class AssociationDetectionCalculatorTest : public ::testing::Test {
+ protected:
+  AssociationDetectionCalculatorTest() {
+    //  0.4                                         ================
+    //                                              |    |    |    |
+    //  0.3 =====================                   |  DET2   |    |
+    //      |    |    |   DET1  |                   |    |   DET4  |
+    //  0.2 |   DET0  |    ===========              ================
+    //      |    |    |    |    |    |
+    //  0.1 =====|===============    |
+    //           |    DET3 |    |    |
+    //  0.0      ================    |
+    //                     |   DET5  |
+    // -0.1                ===========
+    //     0.1  0.2  0.3  0.4  0.5  0.6  0.7  0.8  0.9  1.0  1.1  1.2
+
+    // Detection det_0.
+    det_0 = DetectionWithRelativeLocationData(/*xmin=*/0.1, /*ymin=*/0.1,
+                                              /*width=*/0.2, /*height=*/0.2);
+    det_0.set_detection_id(0);
+
+    // Detection det_1.
+    det_1 = DetectionWithRelativeLocationData(/*xmin=*/0.3, /*ymin=*/0.1,
+                                              /*width=*/0.2, /*height=*/0.2);
+    det_1.set_detection_id(1);
+
+    // Detection det_2.
+    det_2 = DetectionWithRelativeLocationData(/*xmin=*/0.9, /*ymin=*/0.2,
+                                              /*width=*/0.2, /*height=*/0.2);
+    det_2.set_detection_id(2);
+
+    // Detection det_3.
+    det_3 = DetectionWithRelativeLocationData(/*xmin=*/0.2, /*ymin=*/0.0,
+                                              /*width=*/0.3, /*height=*/0.3);
+    det_3.set_detection_id(3);
+
+    // Detection det_4.
+    det_4 = DetectionWithRelativeLocationData(/*xmin=*/1.0, /*ymin=*/0.2,
+                                              /*width=*/0.2, /*height=*/0.2);
+    det_4.set_detection_id(4);
+
+    // Detection det_5.
+    det_5 = DetectionWithRelativeLocationData(/*xmin=*/0.3, /*ymin=*/-0.1,
+                                              /*width=*/0.3, /*height=*/0.3);
+    det_5.set_detection_id(5);
+  }
+
+  ::mediapipe::Detection det_0, det_1, det_2, det_3, det_4, det_5;
+};
+
+TEST_F(AssociationDetectionCalculatorTest, DetectionAssocTest) {
+  CalculatorRunner runner(ParseTextProtoOrDie<CalculatorGraphConfig::Node>(R"(
+    calculator: "AssociationDetectionCalculator"
+    input_stream: "input_vec_0"
+    input_stream: "input_vec_1"
+    input_stream: "input_vec_2"
+    output_stream: "output_vec"
+    options {
+      [mediapipe.AssociationCalculatorOptions.ext] {
+        min_similarity_threshold: 0.1
+      }
+    }
+  )"));
+
+  // Input Stream 0: det_0, det_1, det_2.
+  auto input_vec_0 = absl::make_unique<std::vector<::mediapipe::Detection>>();
+  input_vec_0->push_back(det_0);
+  input_vec_0->push_back(det_1);
+  input_vec_0->push_back(det_2);
+  runner.MutableInputs()->Index(0).packets.push_back(
+      Adopt(input_vec_0.release()).At(Timestamp(1)));
+
+  // Input Stream 1: det_3, det_4.
+  auto input_vec_1 = absl::make_unique<std::vector<::mediapipe::Detection>>();
+  input_vec_1->push_back(det_3);
+  input_vec_1->push_back(det_4);
+  runner.MutableInputs()->Index(1).packets.push_back(
+      Adopt(input_vec_1.release()).At(Timestamp(1)));
+
+  // Input Stream 2: det_5.
+  auto input_vec_2 = absl::make_unique<std::vector<::mediapipe::Detection>>();
+  input_vec_2->push_back(det_5);
+  runner.MutableInputs()->Index(2).packets.push_back(
+      Adopt(input_vec_2.release()).At(Timestamp(1)));
+
+  MP_ASSERT_OK(runner.Run()) << "Calculator execution failed.";
+  const std::vector<Packet>& output = runner.Outputs().Index(0).packets;
+  EXPECT_EQ(1, output.size());
+  const auto& assoc_rects =
+      output[0].Get<std::vector<::mediapipe::Detection>>();
+
+  // det_3 overlaps with det_0, det_1 and det_5 overlaps with det_3. Since det_5
+  // is in the highest priority, we remove other rects. det_4 overlaps with
+  // det_2, and det_4 is higher priority, so we keep it. The final output
+  // therefore contains 2 elements.
+  EXPECT_EQ(2, assoc_rects.size());
+  // Outputs are in order of inputs, so det_4 is before det_5 in output vector.
+
+  // det_4 overlaps with det_2, so new id for det_4 is 2.
+  EXPECT_TRUE(assoc_rects[0].has_detection_id());
+  EXPECT_EQ(2, assoc_rects[0].detection_id());
+  det_4.set_detection_id(2);
+  EXPECT_THAT(assoc_rects[0], EqualsProto(det_4));
+
+  // det_3 overlaps with det_0, so new id for det_3 is 0.
+  // det_3 overlaps with det_1, so new id for det_3 is 1.
+  // det_5 overlaps with det_3, so new id for det_5 is 1.
+  EXPECT_TRUE(assoc_rects[1].has_detection_id());
+  EXPECT_EQ(1, assoc_rects[1].detection_id());
+  det_5.set_detection_id(1);
+  EXPECT_THAT(assoc_rects[1], EqualsProto(det_5));
+}
+
+TEST_F(AssociationDetectionCalculatorTest, DetectionAssocTestWithPrev) {
+  CalculatorRunner runner(ParseTextProtoOrDie<CalculatorGraphConfig::Node>(R"(
+    calculator: "AssociationDetectionCalculator"
+    input_stream: "PREV:input_vec_0"
+    input_stream: "input_vec_1"
+    output_stream: "output_vec"
+    options {
+      [mediapipe.AssociationCalculatorOptions.ext] {
+        min_similarity_threshold: 0.1
+      }
+    }
+  )"));
+
+  // Input Stream 0: det_3, det_4.
+  auto input_vec_0 = absl::make_unique<std::vector<::mediapipe::Detection>>();
+  input_vec_0->push_back(det_3);
+  input_vec_0->push_back(det_4);
+  CollectionItemId prev_input_stream_id =
+      runner.MutableInputs()->GetId("PREV", 0);
+  runner.MutableInputs()
+      ->Get(prev_input_stream_id)
+      .packets.push_back(Adopt(input_vec_0.release()).At(Timestamp(1)));
+
+  // Input Stream 1: det_5.
+  auto input_vec_1 = absl::make_unique<std::vector<::mediapipe::Detection>>();
+  input_vec_1->push_back(det_5);
+  CollectionItemId input_stream_id = runner.MutableInputs()->GetId("", 0);
+  runner.MutableInputs()
+      ->Get(input_stream_id)
+      .packets.push_back(Adopt(input_vec_1.release()).At(Timestamp(1)));
+
+  MP_ASSERT_OK(runner.Run()) << "Calculator execution failed.";
+  const std::vector<Packet>& output = runner.Outputs().Index(0).packets;
+  EXPECT_EQ(1, output.size());
+  const auto& assoc_rects =
+      output[0].Get<std::vector<::mediapipe::Detection>>();
+
+  // det_5 overlaps with det_3 and doesn't overlap with det_4. Since det_4 is
+  // in the PREV input stream, it doesn't get copied to the output, so the final
+  // output contains 1 element.
+  EXPECT_EQ(1, assoc_rects.size());
+
+  // det_5 overlaps with det_3, det_3 is in PREV, so new id for det_5 is 3.
+  EXPECT_TRUE(assoc_rects[0].has_detection_id());
+  EXPECT_EQ(3, assoc_rects[0].detection_id());
+  det_5.set_detection_id(3);
+  EXPECT_THAT(assoc_rects[0], EqualsProto(det_5));
+}
+
+TEST_F(AssociationDetectionCalculatorTest, DetectionAssocTestReverse) {
+  CalculatorRunner runner(ParseTextProtoOrDie<CalculatorGraphConfig::Node>(R"(
+    calculator: "AssociationDetectionCalculator"
+    input_stream: "input_vec_0"
+    input_stream: "input_vec_1"
+    input_stream: "input_vec_2"
+    output_stream: "output_vec"
+    options {
+      [mediapipe.AssociationCalculatorOptions.ext] {
+        min_similarity_threshold: 0.1
+      }
+    }
+  )"));
+
+  // Input Stream 0: det_5.
+  auto input_vec_0 = absl::make_unique<std::vector<::mediapipe::Detection>>();
+  input_vec_0->push_back(det_5);
+  runner.MutableInputs()->Index(0).packets.push_back(
+      Adopt(input_vec_0.release()).At(Timestamp(1)));
+
+  // Input Stream 1: det_3, det_4.
+  auto input_vec_1 = absl::make_unique<std::vector<::mediapipe::Detection>>();
+  input_vec_1->push_back(det_3);
+  input_vec_1->push_back(det_4);
+  runner.MutableInputs()->Index(1).packets.push_back(
+      Adopt(input_vec_1.release()).At(Timestamp(1)));
+
+  // Input Stream 2: det_0, det_1, det_2.
+  auto input_vec_2 = absl::make_unique<std::vector<::mediapipe::Detection>>();
+  input_vec_2->push_back(det_0);
+  input_vec_2->push_back(det_1);
+  input_vec_2->push_back(det_2);
+  runner.MutableInputs()->Index(2).packets.push_back(
+      Adopt(input_vec_2.release()).At(Timestamp(1)));
+
+  MP_ASSERT_OK(runner.Run()) << "Calculator execution failed.";
+  const std::vector<Packet>& output = runner.Outputs().Index(0).packets;
+  EXPECT_EQ(1, output.size());
+  const auto& assoc_rects =
+      output[0].Get<std::vector<::mediapipe::Detection>>();
+
+  // det_3 overlaps with det_5, so det_5 is removed. det_0 overlaps with det_3,
+  // so det_3 is removed as det_0 is in higher priority for keeping. det_2
+  // overlaps with det_4 so det_4 is removed as det_2 is higher priority for
+  // keeping. The final output therefore contains 3 elements.
+  EXPECT_EQ(3, assoc_rects.size());
+  // Outputs are in same order as inputs.
+
+  // det_3 overlaps with det_5, so new id for det_3 is 5.
+  // det_0 overlaps with det_3, so new id for det_0 is 5.
+  EXPECT_TRUE(assoc_rects[0].has_detection_id());
+  EXPECT_EQ(5, assoc_rects[0].detection_id());
+  det_0.set_detection_id(5);
+  EXPECT_THAT(assoc_rects[0], EqualsProto(det_0));
+
+  // det_1 stays with id 1.
+  EXPECT_TRUE(assoc_rects[1].has_detection_id());
+  EXPECT_EQ(1, assoc_rects[1].detection_id());
+  EXPECT_THAT(assoc_rects[1], EqualsProto(det_1));
+
+  // det_2 overlaps with det_4, so new id for det_2 is 4.
+  EXPECT_TRUE(assoc_rects[2].has_detection_id());
+  EXPECT_EQ(4, assoc_rects[2].detection_id());
+  det_2.set_detection_id(4);
+  EXPECT_THAT(assoc_rects[2], EqualsProto(det_2));
+}
+
+class AssociationNormRectCalculatorTest : public ::testing::Test {
+ protected:
+  AssociationNormRectCalculatorTest() {
+    //  0.4                                         ================
+    //                                              |    |    |    |
+    //  0.3 =====================                   |   NR2   |    |
+    //      |    |    |   NR1   |                   |    |    NR4  |
+    //  0.2 |   NR0   |    ===========              ================
+    //      |    |    |    |    |    |
+    //  0.1 =====|===============    |
+    //           |    NR3  |    |    |
+    //  0.0      ================    |
+    //                     |   NR5   |
+    // -0.1                ===========
+    //     0.1  0.2  0.3  0.4  0.5  0.6  0.7  0.8  0.9  1.0  1.1  1.2
+
+    // NormalizedRect nr_0.
+    nr_0.set_x_center(0.2);
+    nr_0.set_y_center(0.2);
+    nr_0.set_width(0.2);
+    nr_0.set_height(0.2);
+
+    // NormalizedRect nr_1.
+    nr_1.set_x_center(0.4);
+    nr_1.set_y_center(0.2);
+    nr_1.set_width(0.2);
+    nr_1.set_height(0.2);
+
+    // NormalizedRect nr_2.
+    nr_2.set_x_center(1.0);
+    nr_2.set_y_center(0.3);
+    nr_2.set_width(0.2);
+    nr_2.set_height(0.2);
+
+    // NormalizedRect nr_3.
+    nr_3.set_x_center(0.35);
+    nr_3.set_y_center(0.15);
+    nr_3.set_width(0.3);
+    nr_3.set_height(0.3);
+
+    // NormalizedRect nr_4.
+    nr_4.set_x_center(1.1);
+    nr_4.set_y_center(0.3);
+    nr_4.set_width(0.2);
+    nr_4.set_height(0.2);
+
+    // NormalizedRect nr_5.
+    nr_5.set_x_center(0.45);
+    nr_5.set_y_center(0.05);
+    nr_5.set_width(0.3);
+    nr_5.set_height(0.3);
+  }
+
+  ::mediapipe::NormalizedRect nr_0, nr_1, nr_2, nr_3, nr_4, nr_5;
+};
+
+TEST_F(AssociationNormRectCalculatorTest, NormRectAssocTest) {
+  CalculatorRunner runner(ParseTextProtoOrDie<CalculatorGraphConfig::Node>(R"(
+    calculator: "AssociationNormRectCalculator"
+    input_stream: "input_vec_0"
+    input_stream: "input_vec_1"
+    input_stream: "input_vec_2"
+    output_stream: "output_vec"
+    options {
+      [mediapipe.AssociationCalculatorOptions.ext] {
+        min_similarity_threshold: 0.1
+      }
+    }
+  )"));
+
+  // Input Stream 0: nr_0, nr_1, nr_2.
+  auto input_vec_0 =
+      absl::make_unique<std::vector<::mediapipe::NormalizedRect>>();
+  input_vec_0->push_back(nr_0);
+  input_vec_0->push_back(nr_1);
+  input_vec_0->push_back(nr_2);
+  runner.MutableInputs()->Index(0).packets.push_back(
+      Adopt(input_vec_0.release()).At(Timestamp(1)));
+
+  // Input Stream 1: nr_3, nr_4.
+  auto input_vec_1 =
+      absl::make_unique<std::vector<::mediapipe::NormalizedRect>>();
+  input_vec_1->push_back(nr_3);
+  input_vec_1->push_back(nr_4);
+  runner.MutableInputs()->Index(1).packets.push_back(
+      Adopt(input_vec_1.release()).At(Timestamp(1)));
+
+  // Input Stream 2: nr_5.
+  auto input_vec_2 =
+      absl::make_unique<std::vector<::mediapipe::NormalizedRect>>();
+  input_vec_2->push_back(nr_5);
+  runner.MutableInputs()->Index(2).packets.push_back(
+      Adopt(input_vec_2.release()).At(Timestamp(1)));
+
+  MP_ASSERT_OK(runner.Run()) << "Calculator execution failed.";
+  const std::vector<Packet>& output = runner.Outputs().Index(0).packets;
+  EXPECT_EQ(1, output.size());
+  const auto& assoc_rects =
+      output[0].Get<std::vector<::mediapipe::NormalizedRect>>();
+
+  // nr_3 overlaps with nr_0, nr_1 and nr_5 overlaps with nr_3. Since nr_5 is
+  // in the highest priority, we remove other rects.
+  // nr_4 overlaps with nr_2, and nr_4 is higher priority, so we keep it.
+  // The final output therefore contains 2 elements.
+  EXPECT_EQ(2, assoc_rects.size());
+  // Outputs are in order of inputs, so nr_4 is before nr_5 in output vector.
+  EXPECT_THAT(assoc_rects[0], EqualsProto(nr_4));
+  EXPECT_THAT(assoc_rects[1], EqualsProto(nr_5));
+}
+
+TEST_F(AssociationNormRectCalculatorTest, NormRectAssocTestReverse) {
+  CalculatorRunner runner(ParseTextProtoOrDie<CalculatorGraphConfig::Node>(R"(
+    calculator: "AssociationNormRectCalculator"
+    input_stream: "input_vec_0"
+    input_stream: "input_vec_1"
+    input_stream: "input_vec_2"
+    output_stream: "output_vec"
+    options {
+      [mediapipe.AssociationCalculatorOptions.ext] {
+        min_similarity_threshold: 0.1
+      }
+    }
+  )"));
+
+  // Input Stream 0: nr_5.
+  auto input_vec_0 =
+      absl::make_unique<std::vector<::mediapipe::NormalizedRect>>();
+  input_vec_0->push_back(nr_5);
+  runner.MutableInputs()->Index(0).packets.push_back(
+      Adopt(input_vec_0.release()).At(Timestamp(1)));
+
+  // Input Stream 1: nr_3, nr_4.
+  auto input_vec_1 =
+      absl::make_unique<std::vector<::mediapipe::NormalizedRect>>();
+  input_vec_1->push_back(nr_3);
+  input_vec_1->push_back(nr_4);
+  runner.MutableInputs()->Index(1).packets.push_back(
+      Adopt(input_vec_1.release()).At(Timestamp(1)));
+
+  // Input Stream 2: nr_0, nr_1, nr_2.
+  auto input_vec_2 =
+      absl::make_unique<std::vector<::mediapipe::NormalizedRect>>();
+  input_vec_2->push_back(nr_0);
+  input_vec_2->push_back(nr_1);
+  input_vec_2->push_back(nr_2);
+  runner.MutableInputs()->Index(2).packets.push_back(
+      Adopt(input_vec_2.release()).At(Timestamp(1)));
+
+  MP_ASSERT_OK(runner.Run()) << "Calculator execution failed.";
+  const std::vector<Packet>& output = runner.Outputs().Index(0).packets;
+  EXPECT_EQ(1, output.size());
+  const auto& assoc_rects =
+      output[0].Get<std::vector<::mediapipe::NormalizedRect>>();
+
+  // nr_3 overlaps with nr_5, so nr_5 is removed. nr_0 overlaps with nr_3, so
+  // nr_3 is removed as nr_0 is in higher priority for keeping. nr_2 overlaps
+  // with nr_4 so nr_4 is removed as nr_2 is higher priority for keeping.
+  // The final output therefore contains 3 elements.
+  EXPECT_EQ(3, assoc_rects.size());
+  // Outputs are in same order as inputs.
+  EXPECT_THAT(assoc_rects[0], EqualsProto(nr_0));
+  EXPECT_THAT(assoc_rects[1], EqualsProto(nr_1));
+  EXPECT_THAT(assoc_rects[2], EqualsProto(nr_2));
+}
+
+TEST_F(AssociationNormRectCalculatorTest, NormRectAssocSingleInputStream) {
+  CalculatorRunner runner(ParseTextProtoOrDie<CalculatorGraphConfig::Node>(R"(
+    calculator: "AssociationNormRectCalculator"
+    input_stream: "input_vec"
+    output_stream: "output_vec"
+    options {
+      [mediapipe.AssociationCalculatorOptions.ext] {
+        min_similarity_threshold: 0.1
+      }
+    }
+  )"));
+
+  // Input Stream : nr_3, nr_5.
+  auto input_vec =
+      absl::make_unique<std::vector<::mediapipe::NormalizedRect>>();
+  input_vec->push_back(nr_3);
+  input_vec->push_back(nr_5);
+  runner.MutableInputs()->Index(0).packets.push_back(
+      Adopt(input_vec.release()).At(Timestamp(1)));
+
+  MP_ASSERT_OK(runner.Run()) << "Calculator execution failed.";
+  const std::vector<Packet>& output = runner.Outputs().Index(0).packets;
+  EXPECT_EQ(1, output.size());
+  const auto& assoc_rects =
+      output[0].Get<std::vector<::mediapipe::NormalizedRect>>();
+
+  // nr_5 overlaps with nr_3. Since nr_5 is after nr_3 in the same input stream
+  // we remove nr_3 and keep nr_5.
+  // The final output therefore contains 1 elements.
+  EXPECT_EQ(1, assoc_rects.size());
+  EXPECT_THAT(assoc_rects[0], EqualsProto(nr_5));
+}
+
+}  // namespace mediapipe

mediapipe/calculators/util/association_detection_calculator.cc

@@ -0,0 +1,77 @@
+// 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/association_calculator.h"
+#include "mediapipe/framework/calculator_context.h"
+#include "mediapipe/framework/calculator_framework.h"
+#include "mediapipe/framework/formats/detection.pb.h"
+#include "mediapipe/framework/formats/location.h"
+#include "mediapipe/framework/port/rectangle.h"
+#include "mediapipe/framework/port/status.h"
+
+namespace mediapipe {
+
+// A subclass of AssociationCalculator<T> for Detection. Example:
+// node {
+//   calculator: "AssociationDetectionCalculator"
+//   input_stream: "PREV:input_vec_0"
+//   input_stream: "input_vec_1"
+//   input_stream: "input_vec_2"
+//   output_stream: "output_vec"
+//   options {
+//     [mediapipe.AssociationCalculatorOptions.ext] {
+//       min_similarity_threshold: 0.1
+//     }
+// }
+class AssociationDetectionCalculator
+    : public AssociationCalculator<::mediapipe::Detection> {
+ public:
+  static ::mediapipe::Status GetContract(CalculatorContract* cc) {
+    return AssociationCalculator<::mediapipe::Detection>::GetContract(cc);
+  }
+
+  ::mediapipe::Status Open(CalculatorContext* cc) override {
+    return AssociationCalculator<::mediapipe::Detection>::Open(cc);
+  }
+
+  ::mediapipe::Status Process(CalculatorContext* cc) override {
+    return AssociationCalculator<::mediapipe::Detection>::Process(cc);
+  }
+
+  ::mediapipe::Status Close(CalculatorContext* cc) override {
+    return AssociationCalculator<::mediapipe::Detection>::Close(cc);
+  }
+
+ protected:
+  ::mediapipe::StatusOr<Rectangle_f> GetRectangle(
+      const ::mediapipe::Detection& input) override {
+    if (!input.has_location_data()) {
+      return ::mediapipe::InternalError("Missing location_data in Detection");
+    }
+    const Location location(input.location_data());
+    return location.GetRelativeBBox();
+  }
+
+  std::pair<bool, int> GetId(const ::mediapipe::Detection& input) override {
+    return {input.has_detection_id(), input.detection_id()};
+  }
+
+  void SetId(::mediapipe::Detection* input, int id) override {
+    input->set_detection_id(id);
+  }
+};
+
+REGISTER_CALCULATOR(AssociationDetectionCalculator);
+
+}  // namespace mediapipe

mediapipe/calculators/util/association_norm_rect_calculator.cc

@@ -0,0 +1,72 @@
+// 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/association_calculator.h"
+#include "mediapipe/framework/calculator_context.h"
+#include "mediapipe/framework/calculator_framework.h"
+#include "mediapipe/framework/formats/rect.pb.h"
+#include "mediapipe/framework/port/rectangle.h"
+#include "mediapipe/framework/port/status.h"
+
+namespace mediapipe {
+
+// A subclass of AssociationCalculator<T> for NormalizedRect. Example use case:
+// node {
+//   calculator: "AssociationNormRectCalculator"
+//   input_stream: "input_vec_0"
+//   input_stream: "input_vec_1"
+//   input_stream: "input_vec_2"
+//   output_stream: "output_vec"
+//   options {
+//     [mediapipe.AssociationCalculatorOptions.ext] {
+//       min_similarity_threshold: 0.1
+//     }
+// }
+class AssociationNormRectCalculator
+    : public AssociationCalculator<::mediapipe::NormalizedRect> {
+ public:
+  static ::mediapipe::Status GetContract(CalculatorContract* cc) {
+    return AssociationCalculator<::mediapipe::NormalizedRect>::GetContract(cc);
+  }
+
+  ::mediapipe::Status Open(CalculatorContext* cc) override {
+    return AssociationCalculator<::mediapipe::NormalizedRect>::Open(cc);
+  }
+
+  ::mediapipe::Status Process(CalculatorContext* cc) override {
+    return AssociationCalculator<::mediapipe::NormalizedRect>::Process(cc);
+  }
+
+  ::mediapipe::Status Close(CalculatorContext* cc) override {
+    return AssociationCalculator<::mediapipe::NormalizedRect>::Close(cc);
+  }
+
+ protected:
+  ::mediapipe::StatusOr<Rectangle_f> GetRectangle(
+      const ::mediapipe::NormalizedRect& input) override {
+    if (!input.has_x_center() || !input.has_y_center() || !input.has_width() ||
+        !input.has_height()) {
+      return ::mediapipe::InternalError(
+          "Missing dimensions in NormalizedRect.");
+    }
+    const float xmin = input.x_center() - input.width() / 2.0;
+    const float ymin = input.y_center() - input.height() / 2.0;
+    // TODO: Support rotation for rectangle.
+    return Rectangle_f(xmin, ymin, input.width(), input.height());
+  }
+};
+
+REGISTER_CALCULATOR(AssociationNormRectCalculator);
+
+}  // namespace mediapipe

mediapipe/calculators/util/collection_has_min_size_calculator.cc

@@ -0,0 +1,26 @@
+
+// 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/collection_has_min_size_calculator.h"
+
+#include "mediapipe/framework/formats/rect.pb.h"
+
+namespace mediapipe {
+
+typedef CollectionHasMinSizeCalculator<std::vector<::mediapipe::NormalizedRect>>
+    NormalizedRectVectorHasMinSizeCalculator;
+REGISTER_CALCULATOR(NormalizedRectVectorHasMinSizeCalculator);
+
+}  // namespace mediapipe

mediapipe/calculators/util/collection_has_min_size_calculator.h

@@ -0,0 +1,84 @@
+// 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.
+
+#ifndef MEDIAPIPE_CALCULATORS_UTIL_COLLECTION_HAS_MIN_SIZE_CALCULATOR_H_
+#define MEDIAPIPE_CALCULATORS_UTIL_COLLECTION_HAS_MIN_SIZE_CALCULATOR_H_
+
+#include <vector>
+
+#include "mediapipe/calculators/util/collection_has_min_size_calculator.pb.h"
+#include "mediapipe/framework/calculator_framework.h"
+#include "mediapipe/framework/port/canonical_errors.h"
+#include "mediapipe/framework/port/ret_check.h"
+#include "mediapipe/framework/port/status.h"
+
+namespace mediapipe {
+
+// Deterimines if an input iterable collection has a minimum size, specified
+// in CollectionHasMinSizeCalculatorOptions. Example usage:
+// node {
+//   calculator: "IntVectorHasMinSizeCalculator"
+//   input_stream: "ITERABLE:input_int_vector"
+//   output_stream: "has_min_ints"
+//  options {
+//    [mediapipe.CollectionHasMinSizeCalculatorOptions.ext] {
+//      min_size: 2
+//    }
+//   }
+// }
+template <typename IterableT>
+class CollectionHasMinSizeCalculator : public CalculatorBase {
+ public:
+  static ::mediapipe::Status GetContract(CalculatorContract* cc) {
+    RET_CHECK(cc->Inputs().HasTag("ITERABLE"));
+    RET_CHECK_EQ(1, cc->Inputs().NumEntries());
+
+    RET_CHECK_EQ(1, cc->Outputs().NumEntries());
+
+    RET_CHECK_GE(
+        cc->Options<::mediapipe::CollectionHasMinSizeCalculatorOptions>()
+            .min_size(),
+        0);
+
+    cc->Inputs().Tag("ITERABLE").Set<IterableT>();
+    cc->Outputs().Index(0).Set<bool>();
+
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Open(CalculatorContext* cc) override {
+    cc->SetOffset(TimestampDiff(0));
+    min_size_ =
+        cc->Options<::mediapipe::CollectionHasMinSizeCalculatorOptions>()
+            .min_size();
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Process(CalculatorContext* cc) override {
+    const IterableT& input = cc->Inputs().Tag("ITERABLE").Get<IterableT>();
+    bool has_min_size = input.size() >= min_size_;
+
+    cc->Outputs().Index(0).AddPacket(
+        MakePacket<bool>(has_min_size).At(cc->InputTimestamp()));
+
+    return ::mediapipe::OkStatus();
+  }
+
+ private:
+  int min_size_ = 0;
+};
+
+}  // namespace mediapipe
+
+#endif  // MEDIAPIPE_CALCULATORS_UTIL_COLLECTION_HAS_MIN_SIZE_CALCULATOR_H_

mediapipe/calculators/util/collection_has_min_size_calculator.proto

@@ -0,0 +1,29 @@
+// 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 CollectionHasMinSizeCalculatorOptions {
+  extend CalculatorOptions {
+    optional CollectionHasMinSizeCalculatorOptions ext = 259397840;
+  }
+
+  // The minimum size an input iterable collection should have for the
+  // calculator to output true.
+  optional int32 min_size = 1 [default = 0];
+}

mediapipe/calculators/util/detection_label_id_to_text_calculator.cc

@@ -19,8 +19,7 @@
 #include "mediapipe/framework/port/status.h"
 #include "mediapipe/util/resource_util.h"
 
-#if defined(MEDIAPIPE_LITE) || defined(__EMSCRIPTEN__) || \
+#if defined(MEDIAPIPE_MOBILE)
-    defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
 #include "mediapipe/util/android/file/base/file.h"
 #include "mediapipe/util/android/file/base/helpers.h"
 #else

mediapipe/calculators/util/filter_collection_calculator.cc

@@ -0,0 +1,34 @@
+
+// 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/filter_collection_calculator.h"
+
+#include <vector>
+
+#include "mediapipe/framework/formats/landmark.pb.h"
+#include "mediapipe/framework/formats/rect.pb.h"
+
+namespace mediapipe {
+
+typedef FilterCollectionCalculator<std::vector<::mediapipe::NormalizedRect>>
+    FilterNormalizedRectCollectionCalculator;
+REGISTER_CALCULATOR(FilterNormalizedRectCollectionCalculator);
+
+typedef FilterCollectionCalculator<
+    std::vector<::mediapipe::NormalizedLandmarkList>>
+    FilterLandmarkListCollectionCalculator;
+REGISTER_CALCULATOR(FilterLandmarkListCollectionCalculator);
+
+}  // namespace mediapipe

mediapipe/calculators/util/filter_collection_calculator.h

@@ -0,0 +1,109 @@
+// 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.
+
+#ifndef MEDIAPIPE_CALCULATORS_UTIL_FILTER_VECTOR_CALCULATOR_H_
+#define MEDIAPIPE_CALCULATORS_UTIL_FILTER_VECTOR_CALCULATOR_H_
+
+#include <vector>
+
+#include "absl/strings/str_cat.h"
+#include "mediapipe/framework/calculator_framework.h"
+#include "mediapipe/framework/port/canonical_errors.h"
+#include "mediapipe/framework/port/ret_check.h"
+#include "mediapipe/framework/port/status.h"
+
+namespace mediapipe {
+
+// A calculator that gates elements of an input collection based on
+// corresponding boolean values of the "CONDITION" vector. If there is no input
+// collection or "CONDITION" vector, the calculator forwards timestamp bounds
+// for downstream calculators. If the "CONDITION" vector has false values for
+// all elements of the input collection, the calculator outputs a packet
+// containing an empty collection.
+// Example usage:
+// node {
+//   calculator: "FilterCollectionCalculator"
+//   input_stream: "ITERABLE:input_collection"
+//   input_stream: "CONDITION:condition_vector"
+//   output_stream: "ITERABLE:output_collection"
+// }
+// This calculator is able to handle collections of copyable types T.
+template <typename IterableT>
+class FilterCollectionCalculator : public CalculatorBase {
+ public:
+  static ::mediapipe::Status GetContract(CalculatorContract* cc) {
+    RET_CHECK(cc->Inputs().HasTag("ITERABLE"));
+    RET_CHECK(cc->Inputs().HasTag("CONDITION"));
+    RET_CHECK(cc->Outputs().HasTag("ITERABLE"));
+
+    cc->Inputs().Tag("ITERABLE").Set<IterableT>();
+    cc->Inputs().Tag("CONDITION").Set<std::vector<bool>>();
+
+    cc->Outputs().Tag("ITERABLE").Set<IterableT>();
+
+    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().Tag("ITERABLE").IsEmpty()) {
+      return ::mediapipe::OkStatus();
+    }
+    if (cc->Inputs().Tag("CONDITION").IsEmpty()) {
+      return ::mediapipe::OkStatus();
+    }
+
+    const std::vector<bool>& filter_by =
+        cc->Inputs().Tag("CONDITION").Get<std::vector<bool>>();
+
+    return FilterCollection<IterableT>(
+        std::is_copy_constructible<typename IterableT::value_type>(), cc,
+        filter_by);
+  }
+
+  template <typename IterableU>
+  ::mediapipe::Status FilterCollection(std::true_type, CalculatorContext* cc,
+                                       const std::vector<bool>& filter_by) {
+    const IterableU& input = cc->Inputs().Tag("ITERABLE").Get<IterableU>();
+    if (input.size() != filter_by.size()) {
+      return ::mediapipe::InternalError(absl::StrCat(
+          "Input vector size: ", input.size(),
+          " doesn't mach condition vector size: ", filter_by.size()));
+    }
+
+    auto output = absl::make_unique<IterableU>();
+    for (int i = 0; i < input.size(); ++i) {
+      if (filter_by[i]) {
+        output->push_back(input[i]);
+      }
+    }
+    cc->Outputs().Tag("ITERABLE").Add(output.release(), cc->InputTimestamp());
+    return ::mediapipe::OkStatus();
+  }
+
+  template <typename IterableU>
+  ::mediapipe::Status FilterCollection(std::false_type, CalculatorContext* cc,
+                                       const std::vector<bool>& filter_by) {
+    return ::mediapipe::InternalError(
+        "Cannot copy input collection to filter it.");
+  }
+};
+
+}  // namespace mediapipe
+
+#endif  // MEDIAPIPE_CALCULATORS_UTIL_FILTER_VECTOR_CALCULATOR_H_

mediapipe/calculators/util/labels_to_render_data_calculator.cc

@@ -93,6 +93,7 @@ REGISTER_CALCULATOR(LabelsToRenderDataCalculator);
 }
 
 ::mediapipe::Status LabelsToRenderDataCalculator::Open(CalculatorContext* cc) {
+  cc->SetOffset(TimestampDiff(0));
   options_ = cc->Options<LabelsToRenderDataCalculatorOptions>();
   num_colors_ = options_.color_size();
   label_height_px_ = std::ceil(options_.font_height_px() * kFontHeightScale);

mediapipe/calculators/util/landmark_letterbox_removal_calculator.cc

@@ -49,7 +49,7 @@ constexpr char kLetterboxPaddingTag[] = "LETTERBOX_PADDING";
 // corresponding input image before letterboxing.
 //
 // Input:
-//   LANDMARKS: An std::vector<NormalizedLandmark> representing landmarks on an
+//   LANDMARKS: A NormalizedLandmarkList representing landmarks on an
 //   letterboxed image.
 //
 //   LETTERBOX_PADDING: An std::array<float, 4> representing the letterbox
@@ -57,7 +57,7 @@ constexpr char kLetterboxPaddingTag[] = "LETTERBOX_PADDING";
 //   image, normalized to [0.f, 1.f] by the letterboxed image dimensions.
 //
 // Output:
-//   LANDMARKS: An std::vector<NormalizedLandmark> representing landmarks with
+//   LANDMARKS: An NormalizedLandmarkList proto representing landmarks with
 //   their locations adjusted to the letterbox-removed (non-padded) image.
 //
 // Usage example:
@@ -74,10 +74,10 @@ class LandmarkLetterboxRemovalCalculator : public CalculatorBase {
               cc->Inputs().HasTag(kLetterboxPaddingTag))
         << "Missing one or more input streams.";
 
-    cc->Inputs().Tag(kLandmarksTag).Set<std::vector<NormalizedLandmark>>();
+    cc->Inputs().Tag(kLandmarksTag).Set<NormalizedLandmarkList>();
     cc->Inputs().Tag(kLetterboxPaddingTag).Set<std::array<float, 4>>();
 
-    cc->Outputs().Tag(kLandmarksTag).Set<std::vector<NormalizedLandmark>>();
+    cc->Outputs().Tag(kLandmarksTag).Set<NormalizedLandmarkList>();
 
     return ::mediapipe::OkStatus();
   }
@@ -94,8 +94,8 @@ class LandmarkLetterboxRemovalCalculator : public CalculatorBase {
       return ::mediapipe::OkStatus();
     }
 
-    const auto& input_landmarks =
+    const NormalizedLandmarkList& input_landmarks =
-        cc->Inputs().Tag(kLandmarksTag).Get<std::vector<NormalizedLandmark>>();
+        cc->Inputs().Tag(kLandmarksTag).Get<NormalizedLandmarkList>();
     const auto& letterbox_padding =
         cc->Inputs().Tag(kLetterboxPaddingTag).Get<std::array<float, 4>>();
 
@@ -104,24 +104,23 @@ class LandmarkLetterboxRemovalCalculator : public CalculatorBase {
     const float left_and_right = letterbox_padding[0] + letterbox_padding[2];
     const float top_and_bottom = letterbox_padding[1] + letterbox_padding[3];
 
-    auto output_landmarks =
+    NormalizedLandmarkList output_landmarks;
-        absl::make_unique<std::vector<NormalizedLandmark>>();
+    for (int i = 0; i < input_landmarks.landmark_size(); ++i) {
-    for (const auto& landmark : input_landmarks) {
+      const NormalizedLandmark& landmark = input_landmarks.landmark(i);
-      NormalizedLandmark new_landmark;
+      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_x(new_x);
-      new_landmark.set_y(new_y);
+      new_landmark->set_y(new_y);
       // Keep z-coord as is.
-      new_landmark.set_z(landmark.z());
+      new_landmark->set_z(landmark.z());
-
-      output_landmarks->emplace_back(new_landmark);
     }
 
     cc->Outputs()
         .Tag(kLandmarksTag)
-        .Add(output_landmarks.release(), cc->InputTimestamp());
+        .AddPacket(MakePacket<NormalizedLandmarkList>(output_landmarks)
+                       .At(cc->InputTimestamp()));
     return ::mediapipe::OkStatus();
   }
 };

mediapipe/calculators/util/landmark_letterbox_removal_calculator_test.cc

@@ -43,10 +43,10 @@ CalculatorGraphConfig::Node GetDefaultNode() {
 TEST(LandmarkLetterboxRemovalCalculatorTest, PaddingLeftRight) {
   CalculatorRunner runner(GetDefaultNode());
 
-  auto landmarks = absl::make_unique<std::vector<NormalizedLandmark>>();
+  auto landmarks = absl::make_unique<NormalizedLandmarkList>();
-  landmarks->push_back(CreateLandmark(0.5f, 0.5f));
+  *landmarks->add_landmark() = CreateLandmark(0.5f, 0.5f);
-  landmarks->push_back(CreateLandmark(0.2f, 0.2f));
+  *landmarks->add_landmark() = CreateLandmark(0.2f, 0.2f);
-  landmarks->push_back(CreateLandmark(0.7f, 0.7f));
+  *landmarks->add_landmark() = CreateLandmark(0.7f, 0.7f);
   runner.MutableInputs()
       ->Tag("LANDMARKS")
       .packets.push_back(
@@ -61,26 +61,28 @@ TEST(LandmarkLetterboxRemovalCalculatorTest, PaddingLeftRight) {
   MP_ASSERT_OK(runner.Run()) << "Calculator execution failed.";
   const std::vector<Packet>& output = runner.Outputs().Tag("LANDMARKS").packets;
   ASSERT_EQ(1, output.size());
-  const auto& output_landmarks =
+  const auto& output_landmarks = output[0].Get<NormalizedLandmarkList>();
-      output[0].Get<std::vector<NormalizedLandmark>>();
 
-  EXPECT_EQ(output_landmarks.size(), 3);
+  EXPECT_EQ(output_landmarks.landmark_size(), 3);
 
-  EXPECT_THAT(output_landmarks[0].x(), testing::FloatNear(0.6f, 1e-5));
+  EXPECT_THAT(output_landmarks.landmark(0).x(), testing::FloatNear(0.6f, 1e-5));
-  EXPECT_THAT(output_landmarks[0].y(), testing::FloatNear(0.5f, 1e-5));
+  EXPECT_THAT(output_landmarks.landmark(0).y(), testing::FloatNear(0.5f, 1e-5));
-  EXPECT_THAT(output_landmarks[1].x(), testing::FloatNear(0.0f, 1e-5));
+  EXPECT_THAT(output_landmarks.landmark(1).x(), testing::FloatNear(0.0f, 1e-5));
-  EXPECT_THAT(output_landmarks[1].y(), testing::FloatNear(0.2f, 1e-5));
+  EXPECT_THAT(output_landmarks.landmark(1).y(), testing::FloatNear(0.2f, 1e-5));
-  EXPECT_THAT(output_landmarks[2].x(), testing::FloatNear(1.0f, 1e-5));
+  EXPECT_THAT(output_landmarks.landmark(2).x(), testing::FloatNear(1.0f, 1e-5));
-  EXPECT_THAT(output_landmarks[2].y(), testing::FloatNear(0.7f, 1e-5));
+  EXPECT_THAT(output_landmarks.landmark(2).y(), testing::FloatNear(0.7f, 1e-5));
 }
 
 TEST(LandmarkLetterboxRemovalCalculatorTest, PaddingTopBottom) {
   CalculatorRunner runner(GetDefaultNode());
 
-  auto landmarks = absl::make_unique<std::vector<NormalizedLandmark>>();
+  auto landmarks = absl::make_unique<NormalizedLandmarkList>();
-  landmarks->push_back(CreateLandmark(0.5f, 0.5f));
+  NormalizedLandmark* landmark = landmarks->add_landmark();
-  landmarks->push_back(CreateLandmark(0.2f, 0.2f));
+  *landmark = CreateLandmark(0.5f, 0.5f);
-  landmarks->push_back(CreateLandmark(0.7f, 0.7f));
+  landmark = landmarks->add_landmark();
+  *landmark = CreateLandmark(0.2f, 0.2f);
+  landmark = landmarks->add_landmark();
+  *landmark = CreateLandmark(0.7f, 0.7f);
   runner.MutableInputs()
       ->Tag("LANDMARKS")
       .packets.push_back(
@@ -95,17 +97,16 @@ TEST(LandmarkLetterboxRemovalCalculatorTest, PaddingTopBottom) {
   MP_ASSERT_OK(runner.Run()) << "Calculator execution failed.";
   const std::vector<Packet>& output = runner.Outputs().Tag("LANDMARKS").packets;
   ASSERT_EQ(1, output.size());
-  const auto& output_landmarks =
+  const auto& output_landmarks = output[0].Get<NormalizedLandmarkList>();
-      output[0].Get<std::vector<NormalizedLandmark>>();
 
-  EXPECT_EQ(output_landmarks.size(), 3);
+  EXPECT_EQ(output_landmarks.landmark_size(), 3);
 
-  EXPECT_THAT(output_landmarks[0].x(), testing::FloatNear(0.5f, 1e-5));
+  EXPECT_THAT(output_landmarks.landmark(0).x(), testing::FloatNear(0.5f, 1e-5));
-  EXPECT_THAT(output_landmarks[0].y(), testing::FloatNear(0.6f, 1e-5));
+  EXPECT_THAT(output_landmarks.landmark(0).y(), testing::FloatNear(0.6f, 1e-5));
-  EXPECT_THAT(output_landmarks[1].x(), testing::FloatNear(0.2f, 1e-5));
+  EXPECT_THAT(output_landmarks.landmark(1).x(), testing::FloatNear(0.2f, 1e-5));
-  EXPECT_THAT(output_landmarks[1].y(), testing::FloatNear(0.0f, 1e-5));
+  EXPECT_THAT(output_landmarks.landmark(1).y(), testing::FloatNear(0.0f, 1e-5));
-  EXPECT_THAT(output_landmarks[2].x(), testing::FloatNear(0.7f, 1e-5));
+  EXPECT_THAT(output_landmarks.landmark(2).x(), testing::FloatNear(0.7f, 1e-5));
-  EXPECT_THAT(output_landmarks[2].y(), testing::FloatNear(1.0f, 1e-5));
+  EXPECT_THAT(output_landmarks.landmark(2).y(), testing::FloatNear(1.0f, 1e-5));
 }
 
 }  // namespace mediapipe

mediapipe/calculators/util/landmark_projection_calculator.cc

@@ -47,13 +47,13 @@ constexpr char kRectTag[] = "NORM_RECT";
 // Projects normalized landmarks in a rectangle to its original coordinates. The
 // rectangle must also be in normalized coordinates.
 // Input:
-//   NORM_LANDMARKS: An std::vector<NormalizedLandmark> representing landmarks
+//   NORM_LANDMARKS: A NormalizedLandmarkList representing landmarks
 //                   in a normalized rectangle.
 //   NORM_RECT: An NormalizedRect representing a normalized rectangle in image
 //              coordinates.
 //
 // Output:
-//   NORM_LANDMARKS: An std::vector<NormalizedLandmark> representing landmarks
+//   NORM_LANDMARKS: A NormalizedLandmarkList representing landmarks
 //                   with their locations adjusted to the image.
 //
 // Usage example:
@@ -70,10 +70,10 @@ class LandmarkProjectionCalculator : public CalculatorBase {
               cc->Inputs().HasTag(kRectTag))
         << "Missing one or more input streams.";
 
-    cc->Inputs().Tag(kLandmarksTag).Set<std::vector<NormalizedLandmark>>();
+    cc->Inputs().Tag(kLandmarksTag).Set<NormalizedLandmarkList>();
     cc->Inputs().Tag(kRectTag).Set<NormalizedRect>();
 
-    cc->Outputs().Tag(kLandmarksTag).Set<std::vector<NormalizedLandmark>>();
+    cc->Outputs().Tag(kLandmarksTag).Set<NormalizedLandmarkList>();
 
     return ::mediapipe::OkStatus();
   }
@@ -92,14 +92,14 @@ class LandmarkProjectionCalculator : public CalculatorBase {
       return ::mediapipe::OkStatus();
     }
 
-    const auto& input_landmarks =
+    const NormalizedLandmarkList& input_landmarks =
-        cc->Inputs().Tag(kLandmarksTag).Get<std::vector<NormalizedLandmark>>();
+        cc->Inputs().Tag(kLandmarksTag).Get<NormalizedLandmarkList>();
     const auto& input_rect = cc->Inputs().Tag(kRectTag).Get<NormalizedRect>();
 
-    auto output_landmarks =
+    NormalizedLandmarkList output_landmarks;
-        absl::make_unique<std::vector<NormalizedLandmark>>();
+    for (int i = 0; i < input_landmarks.landmark_size(); ++i) {
-    for (const auto& landmark : input_landmarks) {
+      const NormalizedLandmark& landmark = input_landmarks.landmark(i);
-      NormalizedLandmark new_landmark;
+      NormalizedLandmark* new_landmark = output_landmarks.add_landmark();
 
       const float x = landmark.x() - 0.5f;
       const float y = landmark.y() - 0.5f;
@@ -110,17 +110,16 @@ class LandmarkProjectionCalculator : public CalculatorBase {
       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_x(new_x);
-      new_landmark.set_y(new_y);
+      new_landmark->set_y(new_y);
       // Keep z-coord as is.
-      new_landmark.set_z(landmark.z());
+      new_landmark->set_z(landmark.z());
-
-      output_landmarks->emplace_back(new_landmark);
     }
 
     cc->Outputs()
         .Tag(kLandmarksTag)
-        .Add(output_landmarks.release(), cc->InputTimestamp());
+        .AddPacket(MakePacket<NormalizedLandmarkList>(output_landmarks)
+                       .At(cc->InputTimestamp()));
     return ::mediapipe::OkStatus();
   }
 };

mediapipe/calculators/util/landmarks_to_detection_calculator.cc

@@ -28,8 +28,7 @@ namespace {
 constexpr char kDetectionTag[] = "DETECTION";
 constexpr char kNormalizedLandmarksTag[] = "NORM_LANDMARKS";
 
-Detection ConvertLandmarksToDetection(
+Detection ConvertLandmarksToDetection(const NormalizedLandmarkList& landmarks) {
-    const std::vector<NormalizedLandmark>& landmarks) {
   Detection detection;
   LocationData* location_data = detection.mutable_location_data();
 
@@ -37,7 +36,8 @@ Detection ConvertLandmarksToDetection(
   float x_max = std::numeric_limits<float>::min();
   float y_min = std::numeric_limits<float>::max();
   float y_max = std::numeric_limits<float>::min();
-  for (const auto& landmark : landmarks) {
+  for (int i = 0; i < landmarks.landmark_size(); ++i) {
+    const NormalizedLandmark& landmark = landmarks.landmark(i);
     x_min = std::min(x_min, landmark.x());
     x_max = std::max(x_max, landmark.x());
     y_min = std::min(y_min, landmark.y());
@@ -67,7 +67,7 @@ Detection ConvertLandmarksToDetection(
 // to specify a subset of landmarks for creating the detection.
 //
 // Input:
-//  NOMR_LANDMARKS: A vector of NormalizedLandmark.
+//  NOMR_LANDMARKS: A NormalizedLandmarkList proto.
 //
 // Output:
 //   DETECTION: A Detection proto.
@@ -95,9 +95,7 @@ REGISTER_CALCULATOR(LandmarksToDetectionCalculator);
   RET_CHECK(cc->Inputs().HasTag(kNormalizedLandmarksTag));
   RET_CHECK(cc->Outputs().HasTag(kDetectionTag));
   // TODO: Also support converting Landmark to Detection.
-  cc->Inputs()
+  cc->Inputs().Tag(kNormalizedLandmarksTag).Set<NormalizedLandmarkList>();
-      .Tag(kNormalizedLandmarksTag)
-      .Set<std::vector<NormalizedLandmark>>();
   cc->Outputs().Tag(kDetectionTag).Set<Detection>();
 
   return ::mediapipe::OkStatus();
@@ -113,19 +111,20 @@ REGISTER_CALCULATOR(LandmarksToDetectionCalculator);
 
 ::mediapipe::Status LandmarksToDetectionCalculator::Process(
     CalculatorContext* cc) {
-  const auto& landmarks = cc->Inputs()
+  const auto& landmarks =
-                              .Tag(kNormalizedLandmarksTag)
+      cc->Inputs().Tag(kNormalizedLandmarksTag).Get<NormalizedLandmarkList>();
-                              .Get<std::vector<NormalizedLandmark>>();
+  RET_CHECK_GT(landmarks.landmark_size(), 0)
-  RET_CHECK_GT(landmarks.size(), 0) << "Input landmark vector is empty.";
+      << "Input landmark vector is empty.";
 
   auto detection = absl::make_unique<Detection>();
   if (options_.selected_landmark_indices_size()) {
-    std::vector<NormalizedLandmark> subset_landmarks(
+    NormalizedLandmarkList subset_landmarks;
-        options_.selected_landmark_indices_size());
+    for (int i = 0; i < options_.selected_landmark_indices_size(); ++i) {
-    for (int i = 0; i < subset_landmarks.size(); ++i) {
+      RET_CHECK_LT(options_.selected_landmark_indices(i),
-      RET_CHECK_LT(options_.selected_landmark_indices(i), landmarks.size())
+                   landmarks.landmark_size())
           << "Index of landmark subset is out of range.";
-      subset_landmarks[i] = landmarks[options_.selected_landmark_indices(i)];
+      *subset_landmarks.add_landmark() =
+          landmarks.landmark(options_.selected_landmark_indices(i));
     }
     *detection = ConvertLandmarksToDetection(subset_landmarks);
   } else {

mediapipe/calculators/util/landmarks_to_floats_calculator.cc

@@ -48,7 +48,7 @@ constexpr char kMatrixTag[] = "MATRIX";
 
 // Converts a vector of landmarks to a vector of floats or a matrix.
 // Input:
-//   NORM_LANDMARKS: An std::vector<NormalizedLandmark>.
+//   NORM_LANDMARKS: A NormalizedLandmarkList proto.
 //
 // Output:
 //   FLOATS(optional): A vector of floats from flattened landmarks.
@@ -63,7 +63,7 @@ constexpr char kMatrixTag[] = "MATRIX";
 class LandmarksToFloatsCalculator : public CalculatorBase {
  public:
   static ::mediapipe::Status GetContract(CalculatorContract* cc) {
-    cc->Inputs().Tag(kLandmarksTag).Set<std::vector<NormalizedLandmark>>();
+    cc->Inputs().Tag(kLandmarksTag).Set<NormalizedLandmarkList>();
     RET_CHECK(cc->Outputs().HasTag(kFloatsTag) ||
               cc->Outputs().HasTag(kMatrixTag));
     if (cc->Outputs().HasTag(kFloatsTag)) {
@@ -94,11 +94,12 @@ class LandmarksToFloatsCalculator : public CalculatorBase {
     }
 
     const auto& input_landmarks =
-        cc->Inputs().Tag(kLandmarksTag).Get<std::vector<NormalizedLandmark>>();
+        cc->Inputs().Tag(kLandmarksTag).Get<NormalizedLandmarkList>();
 
     if (cc->Outputs().HasTag(kFloatsTag)) {
       auto output_floats = absl::make_unique<std::vector<float>>();
-      for (const auto& landmark : input_landmarks) {
+      for (int i = 0; i < input_landmarks.landmark_size(); ++i) {
+        const NormalizedLandmark& landmark = input_landmarks.landmark(i);
         output_floats->emplace_back(landmark.x());
         if (num_dimensions_ > 1) {
           output_floats->emplace_back(landmark.y());
@@ -113,14 +114,14 @@ class LandmarksToFloatsCalculator : public CalculatorBase {
           .Add(output_floats.release(), cc->InputTimestamp());
     } else {
       auto output_matrix = absl::make_unique<Matrix>();
-      output_matrix->setZero(num_dimensions_, input_landmarks.size());
+      output_matrix->setZero(num_dimensions_, input_landmarks.landmark_size());
-      for (int i = 0; i < input_landmarks.size(); ++i) {
+      for (int i = 0; i < input_landmarks.landmark_size(); ++i) {
-        (*output_matrix)(0, i) = input_landmarks[i].x();
+        (*output_matrix)(0, i) = input_landmarks.landmark(i).x();
         if (num_dimensions_ > 1) {
-          (*output_matrix)(1, i) = input_landmarks[i].y();
+          (*output_matrix)(1, i) = input_landmarks.landmark(i).y();
         }
         if (num_dimensions_ > 2) {
-          (*output_matrix)(2, i) = input_landmarks[i].z();
+          (*output_matrix)(2, i) = input_landmarks.landmark(i).z();
         }
       }
       cc->Outputs()

mediapipe/calculators/util/landmarks_to_render_data_calculator.cc

@@ -46,12 +46,13 @@ inline float Remap(float x, float lo, float hi, float scale) {
   return (x - lo) / (hi - lo + 1e-6) * scale;
 }
 
-template <class LandmarkType>
+template <class LandmarkListType, class LandmarkType>
-inline void GetMinMaxZ(const std::vector<LandmarkType>& landmarks, float* z_min,
+inline void GetMinMaxZ(const LandmarkListType& landmarks, float* z_min,
                        float* z_max) {
   *z_min = std::numeric_limits<float>::max();
   *z_max = std::numeric_limits<float>::min();
-  for (const auto& landmark : landmarks) {
+  for (int i = 0; i < landmarks.landmark_size(); ++i) {
+    const LandmarkType& landmark = landmarks.landmark(i);
     *z_min = std::min(landmark.z(), *z_min);
     *z_max = std::max(landmark.z(), *z_max);
   }
@@ -73,7 +74,7 @@ void SetColorSizeValueFromZ(float z, float z_min, float z_max,
 }  // namespace
 
 // A calculator that converts Landmark proto to RenderData proto for
-// visualization. The input should be std::vector<Landmark>. It is also possible
+// visualization. The input should be LandmarkList proto. It is also possible
 // to specify the connections between landmarks.
 //
 // Example config:
@@ -121,11 +122,11 @@ class LandmarksToRenderDataCalculator : public CalculatorBase {
       const LandmarksToRenderDataCalculatorOptions& options, bool normalized,
       int gray_val1, int gray_val2, RenderData* render_data);
 
-  template <class LandmarkType>
+  template <class LandmarkListType>
-  void AddConnections(const std::vector<LandmarkType>& landmarks,
+  void AddConnections(const LandmarkListType& landmarks, bool normalized,
-                      bool normalized, RenderData* render_data);
+                      RenderData* render_data);
-  template <class LandmarkType>
+  template <class LandmarkListType>
-  void AddConnectionsWithDepth(const std::vector<LandmarkType>& landmarks,
+  void AddConnectionsWithDepth(const LandmarkListType& landmarks,
                                bool normalized, float min_z, float max_z,
                                RenderData* render_data);
 
@@ -144,10 +145,10 @@ REGISTER_CALCULATOR(LandmarksToRenderDataCalculator);
          "normalized landmarks.";
 
   if (cc->Inputs().HasTag(kLandmarksTag)) {
-    cc->Inputs().Tag(kLandmarksTag).Set<std::vector<Landmark>>();
+    cc->Inputs().Tag(kLandmarksTag).Set<LandmarkList>();
   }
   if (cc->Inputs().HasTag(kNormLandmarksTag)) {
-    cc->Inputs().Tag(kNormLandmarksTag).Set<std::vector<NormalizedLandmark>>();
+    cc->Inputs().Tag(kNormLandmarksTag).Set<NormalizedLandmarkList>();
   }
   cc->Outputs().Tag(kRenderDataTag).Set<RenderData>();
   return ::mediapipe::OkStatus();
@@ -169,16 +170,17 @@ REGISTER_CALCULATOR(LandmarksToRenderDataCalculator);
   float z_max = 0.f;
 
   if (cc->Inputs().HasTag(kLandmarksTag)) {
-    const auto& landmarks =
+    const LandmarkList& landmarks =
-        cc->Inputs().Tag(kLandmarksTag).Get<std::vector<Landmark>>();
+        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(landmarks, &z_min, &z_max);
+      GetMinMaxZ<LandmarkList, Landmark>(landmarks, &z_min, &z_max);
     }
     // Only change rendering if there are actually z values other than 0.
     visualize_depth &= ((z_max - z_min) > 1e-3);
-    for (const auto& landmark : landmarks) {
+    for (int i = 0; i < landmarks.landmark_size(); ++i) {
+      const Landmark& landmark = landmarks.landmark(i);
       auto* landmark_data_render =
           AddPointRenderData(options_, render_data.get());
       if (visualize_depth) {
@@ -191,25 +193,27 @@ REGISTER_CALCULATOR(LandmarksToRenderDataCalculator);
       landmark_data->set_y(landmark.y());
     }
     if (visualize_depth) {
-      AddConnectionsWithDepth(landmarks, /*normalized=*/false, z_min, z_max,
+      AddConnectionsWithDepth<LandmarkList>(landmarks, /*normalized=*/false,
-                              render_data.get());
+                                            z_min, z_max, render_data.get());
     } else {
-      AddConnections(landmarks, /*normalized=*/false, render_data.get());
+      AddConnections<LandmarkList>(landmarks, /*normalized=*/false,
+                                   render_data.get());
     }
   }
 
   if (cc->Inputs().HasTag(kNormLandmarksTag)) {
-    const auto& landmarks = cc->Inputs()
+    const NormalizedLandmarkList& landmarks =
-                                .Tag(kNormLandmarksTag)
+        cc->Inputs().Tag(kNormLandmarksTag).Get<NormalizedLandmarkList>();
-                                .Get<std::vector<NormalizedLandmark>>();
     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(landmarks, &z_min, &z_max);
+      GetMinMaxZ<NormalizedLandmarkList, NormalizedLandmark>(landmarks, &z_min,
+                                                             &z_max);
     }
     // Only change rendering if there are actually z values other than 0.
     visualize_depth &= ((z_max - z_min) > 1e-3);
-    for (const auto& landmark : landmarks) {
+    for (int i = 0; i < landmarks.landmark_size(); ++i) {
+      const NormalizedLandmark& landmark = landmarks.landmark(i);
       auto* landmark_data_render =
           AddPointRenderData(options_, render_data.get());
       if (visualize_depth) {
@@ -222,10 +226,11 @@ REGISTER_CALCULATOR(LandmarksToRenderDataCalculator);
       landmark_data->set_y(landmark.y());
     }
     if (visualize_depth) {
-      AddConnectionsWithDepth(landmarks, /*normalized=*/true, z_min, z_max,
+      AddConnectionsWithDepth<NormalizedLandmarkList>(
-                              render_data.get());
+          landmarks, /*normalized=*/true, z_min, z_max, render_data.get());
     } else {
-      AddConnections(landmarks, /*normalized=*/true, render_data.get());
+      AddConnections<NormalizedLandmarkList>(landmarks, /*normalized=*/true,
+                                             render_data.get());
     }
   }
 
@@ -235,13 +240,13 @@ REGISTER_CALCULATOR(LandmarksToRenderDataCalculator);
   return ::mediapipe::OkStatus();
 }
 
-template <class LandmarkType>
+template <class LandmarkListType>
 void LandmarksToRenderDataCalculator::AddConnectionsWithDepth(
-    const std::vector<LandmarkType>& landmarks, bool normalized, float min_z,
+    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[options_.landmark_connections(i)];
+    const auto& ld0 = landmarks.landmark(options_.landmark_connections(i));
-    const auto& ld1 = landmarks[options_.landmark_connections(i + 1)];
+    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 =
@@ -272,13 +277,13 @@ void LandmarksToRenderDataCalculator::AddConnectionToRenderData(
   connection_annotation->set_thickness(options.thickness());
 }
 
-template <class LandmarkType>
+template <class LandmarkListType>
 void LandmarksToRenderDataCalculator::AddConnections(
-    const std::vector<LandmarkType>& landmarks, bool normalized,
+    const LandmarkListType& landmarks, bool normalized,
     RenderData* render_data) {
   for (int i = 0; i < options_.landmark_connections_size(); i += 2) {
-    const auto& ld0 = landmarks[options_.landmark_connections(i)];
+    const auto& ld0 = landmarks.landmark(options_.landmark_connections(i));
-    const auto& ld1 = landmarks[options_.landmark_connections(i + 1)];
+    const auto& ld1 = landmarks.landmark(options_.landmark_connections(i + 1));
     AddConnectionToRenderData(ld0.x(), ld0.y(), ld1.x(), ld1.y(), options_,
                               normalized, render_data);
   }

mediapipe/calculators/util/rect_transformation_calculator.cc

@@ -23,7 +23,9 @@ namespace mediapipe {
 namespace {
 
 constexpr char kNormRectTag[] = "NORM_RECT";
+constexpr char kNormRectsTag[] = "NORM_RECTS";
 constexpr char kRectTag[] = "RECT";
+constexpr char kRectsTag[] = "RECTS";
 constexpr char kImageSizeTag[] = "IMAGE_SIZE";
 
 // Wraps around an angle in radians to within -M_PI and M_PI.
@@ -72,17 +74,31 @@ REGISTER_CALCULATOR(RectTransformationCalculator);
 
 ::mediapipe::Status RectTransformationCalculator::GetContract(
     CalculatorContract* cc) {
-  RET_CHECK(cc->Inputs().HasTag(kNormRectTag) ^ cc->Inputs().HasTag(kRectTag));
+  RET_CHECK_EQ((cc->Inputs().HasTag(kNormRectTag) ? 1 : 0) +
+                   (cc->Inputs().HasTag(kNormRectsTag) ? 1 : 0) +
+                   (cc->Inputs().HasTag(kRectTag) ? 1 : 0) +
+                   (cc->Inputs().HasTag(kRectsTag) ? 1 : 0),
+               1);
   if (cc->Inputs().HasTag(kRectTag)) {
     cc->Inputs().Tag(kRectTag).Set<Rect>();
     cc->Outputs().Index(0).Set<Rect>();
   }
+  if (cc->Inputs().HasTag(kRectsTag)) {
+    cc->Inputs().Tag(kRectsTag).Set<std::vector<Rect>>();
+    cc->Outputs().Index(0).Set<std::vector<Rect>>();
+  }
   if (cc->Inputs().HasTag(kNormRectTag)) {
     RET_CHECK(cc->Inputs().HasTag(kImageSizeTag));
     cc->Inputs().Tag(kNormRectTag).Set<NormalizedRect>();
     cc->Inputs().Tag(kImageSizeTag).Set<std::pair<int, int>>();
     cc->Outputs().Index(0).Set<NormalizedRect>();
   }
+  if (cc->Inputs().HasTag(kNormRectsTag)) {
+    RET_CHECK(cc->Inputs().HasTag(kImageSizeTag));
+    cc->Inputs().Tag(kNormRectsTag).Set<std::vector<NormalizedRect>>();
+    cc->Inputs().Tag(kImageSizeTag).Set<std::pair<int, int>>();
+    cc->Outputs().Index(0).Set<std::vector<NormalizedRect>>();
+  }
 
   return ::mediapipe::OkStatus();
 }
@@ -105,7 +121,17 @@ REGISTER_CALCULATOR(RectTransformationCalculator);
     cc->Outputs().Index(0).AddPacket(
         MakePacket<Rect>(rect).At(cc->InputTimestamp()));
   }
-
+  if (cc->Inputs().HasTag(kRectsTag) &&
+      !cc->Inputs().Tag(kRectsTag).IsEmpty()) {
+    auto rects = cc->Inputs().Tag(kRectsTag).Get<std::vector<Rect>>();
+    auto output_rects = absl::make_unique<std::vector<Rect>>(rects.size());
+    for (int i = 0; i < rects.size(); ++i) {
+      output_rects->at(i) = rects[i];
+      auto it = output_rects->begin() + i;
+      TransformRect(&(*it));
+    }
+    cc->Outputs().Index(0).Add(output_rects.release(), cc->InputTimestamp());
+  }
   if (cc->Inputs().HasTag(kNormRectTag) &&
       !cc->Inputs().Tag(kNormRectTag).IsEmpty()) {
     auto rect = cc->Inputs().Tag(kNormRectTag).Get<NormalizedRect>();
@@ -115,6 +141,21 @@ REGISTER_CALCULATOR(RectTransformationCalculator);
     cc->Outputs().Index(0).AddPacket(
         MakePacket<NormalizedRect>(rect).At(cc->InputTimestamp()));
   }
+  if (cc->Inputs().HasTag(kNormRectsTag) &&
+      !cc->Inputs().Tag(kNormRectsTag).IsEmpty()) {
+    auto rects =
+        cc->Inputs().Tag(kNormRectsTag).Get<std::vector<NormalizedRect>>();
+    const auto& image_size =
+        cc->Inputs().Tag(kImageSizeTag).Get<std::pair<int, int>>();
+    auto output_rects =
+        absl::make_unique<std::vector<NormalizedRect>>(rects.size());
+    for (int i = 0; i < rects.size(); ++i) {
+      output_rects->at(i) = rects[i];
+      auto it = output_rects->begin() + i;
+      TransformNormalizedRect(&(*it), image_size.first, image_size.second);
+    }
+    cc->Outputs().Index(0).Add(output_rects.release(), cc->InputTimestamp());
+  }
 
   return ::mediapipe::OkStatus();
 }

mediapipe/calculators/util/top_k_scores_calculator.cc

@@ -29,8 +29,7 @@
 #include "mediapipe/framework/port/statusor.h"
 #include "mediapipe/util/resource_util.h"
 
-#if defined(MEDIAPIPE_LITE) || defined(__EMSCRIPTEN__) || \
+#if defined(MEDIAPIPE_MOBILE)
-    defined(__ANDROID__) || (defined(__APPLE__) && !TARGET_OS_OSX)
 #include "mediapipe/util/android/file/base/file.h"
 #include "mediapipe/util/android/file/base/helpers.h"
 #else

mediapipe/calculators/video/BUILD

@@ -13,12 +13,16 @@
 # limitations under the License.
 #
 
+load("//mediapipe/framework/port:build_config.bzl", "mediapipe_cc_proto_library")
+load(
+    "//mediapipe/framework/tool:mediapipe_graph.bzl",
+    "mediapipe_binary_graph",
+)
+
 licenses(["notice"])  # Apache 2.0
 
 package(default_visibility = ["//visibility:private"])
 
-load("//mediapipe/framework/port:build_config.bzl", "mediapipe_cc_proto_library")
-
 proto_library(
     name = "flow_to_image_calculator_proto",
     srcs = ["flow_to_image_calculator.proto"],
@@ -52,9 +56,7 @@ mediapipe_cc_proto_library(
 cc_library(
     name = "flow_to_image_calculator",
     srcs = ["flow_to_image_calculator.cc"],
-    visibility = [
+    visibility = ["//visibility:public"],
-        "//visibility:public",
-    ],
     deps = [
         "//mediapipe/calculators/video:flow_to_image_calculator_cc_proto",
         "//mediapipe/calculators/video/tool:flow_quantizer_model",
@@ -129,10 +131,20 @@ cc_library(
     alwayslink = 1,
 )
 
+filegroup(
+    name = "test_videos",
+    srcs = [
+        "testdata/format_FLV_H264_AAC.video",
+        "testdata/format_MKV_VP8_VORBIS.video",
+        "testdata/format_MP4_AVC720P_AAC.video",
+    ],
+    visibility = ["//visibility:public"],
+)
+
 cc_test(
     name = "opencv_video_decoder_calculator_test",
     srcs = ["opencv_video_decoder_calculator_test.cc"],
-    data = ["//mediapipe/calculators/video/testdata:test_videos"],
+    data = [":test_videos"],
     deps = [
         ":opencv_video_decoder_calculator",
         "//mediapipe/framework:calculator_runner",
@@ -151,7 +163,7 @@ cc_test(
 cc_test(
     name = "opencv_video_encoder_calculator_test",
     srcs = ["opencv_video_encoder_calculator_test.cc"],
-    data = ["//mediapipe/calculators/video/testdata:test_videos"],
+    data = [":test_videos"],
     deps = [
         ":opencv_video_decoder_calculator",
         ":opencv_video_encoder_calculator",
@@ -175,7 +187,6 @@ cc_test(
 cc_test(
     name = "tvl1_optical_flow_calculator_test",
     srcs = ["tvl1_optical_flow_calculator_test.cc"],
-    data = ["//mediapipe/calculators/image/testdata:test_images"],
     deps = [
         ":tvl1_optical_flow_calculator",
         "//mediapipe/framework:calculator_framework",

mediapipe/calculators/video/opencv_video_decoder_calculator.cc

@@ -12,6 +12,8 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+#include <stdlib.h>
+
 #include "mediapipe/framework/calculator_framework.h"
 #include "mediapipe/framework/formats/image_format.pb.h"
 #include "mediapipe/framework/formats/image_frame.h"
@@ -66,6 +68,20 @@ ImageFormat::Format GetImageFormat(int num_channels) {
 //   output_stream: "VIDEO:video_frames"
 //   output_stream: "VIDEO_PRESTREAM:video_header"
 // }
+//
+// OpenCV's VideoCapture doesn't decode audio tracks. If the audio tracks need
+// to be saved, specify an output side packet with tag "SAVED_AUDIO_PATH".
+// The calculator will call FFmpeg binary to save audio tracks as an aac file.
+//
+// Example config:
+// node {
+//   calculator: "OpenCvVideoDecoderCalculator"
+//   input_side_packet: "INPUT_FILE_PATH:input_file_path"
+//   output_side_packet: "SAVED_AUDIO_PATH:audio_path"
+//   output_stream: "VIDEO:video_frames"
+//   output_stream: "VIDEO_PRESTREAM:video_header"
+// }
+//
 class OpenCvVideoDecoderCalculator : public CalculatorBase {
  public:
   static ::mediapipe::Status GetContract(CalculatorContract* cc) {
@@ -74,6 +90,9 @@ class OpenCvVideoDecoderCalculator : public CalculatorBase {
     if (cc->Outputs().HasTag("VIDEO_PRESTREAM")) {
       cc->Outputs().Tag("VIDEO_PRESTREAM").Set<VideoHeader>();
     }
+    if (cc->OutputSidePackets().HasTag("SAVED_AUDIO_PATH")) {
+      cc->OutputSidePackets().Tag("SAVED_AUDIO_PATH").Set<std::string>();
+    }
     return ::mediapipe::OkStatus();
   }
 
@@ -123,9 +142,29 @@ class OpenCvVideoDecoderCalculator : public CalculatorBase {
       cc->Outputs()
           .Tag("VIDEO_PRESTREAM")
           .Add(header.release(), Timestamp::PreStream());
+      cc->Outputs().Tag("VIDEO_PRESTREAM").Close();
     }
     // Rewind to the very first frame.
     cap_->set(cv::CAP_PROP_POS_AVI_RATIO, 0);
+
+    if (cc->OutputSidePackets().HasTag("SAVED_AUDIO_PATH")) {
+#ifdef HAVE_FFMPEG
+      std::string saved_audio_path = std::tmpnam(nullptr);
+      system(absl::StrCat("ffmpeg -nostats -loglevel 0 -i ", input_file_path,
+                          " -vn -f adts ", saved_audio_path)
+                 .c_str());
+      cc->OutputSidePackets()
+          .Tag("SAVED_AUDIO_PATH")
+          .Set(MakePacket<std::string>(saved_audio_path));
+
+#else
+      return ::mediapipe::InvalidArgumentErrorBuilder(MEDIAPIPE_LOC)
+             << "OpenCVVideoDecoderCalculator can't save the audio file "
+                "because FFmpeg is not installed. Please remove "
+                "output_side_packet: \"SAVED_AUDIO_PATH\" from the node "
+                "config.";
+#endif
+    }
     return ::mediapipe::OkStatus();
   }
 

mediapipe/calculators/video/opencv_video_encoder_calculator.cc

@@ -12,6 +12,8 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+#include <stdlib.h>
+
 #include <memory>
 #include <string>
 #include <vector>
@@ -39,8 +41,7 @@ namespace mediapipe {
 // packet. Currently, the calculator only supports one video stream (in
 // mediapipe::ImageFrame).
 //
-// Example config to generate the output video file:
+// Example config:
-//
 // node {
 //   calculator: "OpenCvVideoEncoderCalculator"
 //   input_stream: "VIDEO:video"
@@ -53,6 +54,26 @@ namespace mediapipe {
 //     }
 //   }
 // }
+//
+// OpenCV's VideoWriter doesn't encode audio. If an input side packet with tag
+// "AUDIO_FILE_PATH" is specified, the calculator will call FFmpeg binary to
+// attach the audio file to the video as the last step in Close().
+//
+// Example config:
+// node {
+//   calculator: "OpenCvVideoEncoderCalculator"
+//   input_stream: "VIDEO:video"
+//   input_stream: "VIDEO_PRESTREAM:video_header"
+//   input_side_packet: "OUTPUT_FILE_PATH:output_file_path"
+//   input_side_packet: "AUDIO_FILE_PATH:audio_path"
+//   node_options {
+//     [type.googleapis.com/mediapipe.OpenCvVideoEncoderCalculatorOptions]: {
+//        codec: "avc1"
+//        video_format: "mp4"
+//     }
+//   }
+// }
+//
 class OpenCvVideoEncoderCalculator : public CalculatorBase {
  public:
   static ::mediapipe::Status GetContract(CalculatorContract* cc);
@@ -77,6 +98,9 @@ class OpenCvVideoEncoderCalculator : public CalculatorBase {
   }
   RET_CHECK(cc->InputSidePackets().HasTag("OUTPUT_FILE_PATH"));
   cc->InputSidePackets().Tag("OUTPUT_FILE_PATH").Set<std::string>();
+  if (cc->InputSidePackets().HasTag("AUDIO_FILE_PATH")) {
+    cc->InputSidePackets().Tag("AUDIO_FILE_PATH").Set<std::string>();
+  }
   return ::mediapipe::OkStatus();
 }
 
@@ -155,6 +179,27 @@ class OpenCvVideoEncoderCalculator : public CalculatorBase {
   if (writer_ && writer_->isOpened()) {
     writer_->release();
   }
+  if (cc->InputSidePackets().HasTag("AUDIO_FILE_PATH")) {
+#ifdef HAVE_FFMPEG
+    const std::string& audio_file_path =
+        cc->InputSidePackets().Tag("AUDIO_FILE_PATH").Get<std::string>();
+    // A temp output file is needed because FFmpeg can't do in-place editing.
+    const std::string temp_file_path = std::tmpnam(nullptr);
+    system(absl::StrCat("mv ", output_file_path_, " ", temp_file_path,
+                        "&& ffmpeg -nostats -loglevel 0 -i ", temp_file_path,
+                        " -i ", audio_file_path,
+                        "  -c copy -map 0:v:0 -map 1:a:0 ", output_file_path_,
+                        "&& rm ", temp_file_path)
+               .c_str());
+
+#else
+    return ::mediapipe::InvalidArgumentErrorBuilder(MEDIAPIPE_LOC)
+           << "OpenCVVideoEncoderCalculator can't attach the audio tracks to "
+              "the video because FFmpeg is not installed. Please remove "
+              "input_side_packet: \"AUDIO_FILE_PATH\" from the node "
+              "config.";
+#endif
+  }
   return ::mediapipe::OkStatus();
 }
 

mediapipe/calculators/video/testdata/BUILD

@@ -1,26 +0,0 @@
-# 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.
-#
-
-licenses(["notice"])  # Apache 2.0
-
-filegroup(
-    name = "test_videos",
-    srcs = [
-        "format_FLV_H264_AAC.video",
-        "format_MKV_VP8_VORBIS.video",
-        "format_MP4_AVC720P_AAC.video",
-    ],
-    visibility = ["//visibility:public"],
-)

mediapipe/calculators/video/tool/BUILD

@@ -13,24 +13,24 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+load("//mediapipe/framework/port:build_config.bzl", "mediapipe_cc_proto_library")
+
 licenses(["notice"])  # Apache 2.0
 
 package(default_visibility = ["//mediapipe/calculators/video:__subpackages__"])
 
 exports_files(["LICENSE"])
 
-load("//mediapipe/framework/port:build_config.bzl", "mediapipe_cc_proto_library")
-
 proto_library(
     name = "flow_quantizer_model_proto",
     srcs = ["flow_quantizer_model.proto"],
-    visibility = ["//mediapipe:__subpackages__"],
+    visibility = ["//visibility:public"],
 )
 
 mediapipe_cc_proto_library(
     name = "flow_quantizer_model_cc_proto",
     srcs = ["flow_quantizer_model.proto"],
-    visibility = ["//mediapipe:__subpackages__"],
+    visibility = ["//visibility:public"],
     deps = [":flow_quantizer_model_proto"],
 )
 

mediapipe/docs/android_archive_library.md

@@ -92,7 +92,7 @@ project.
     MediaPipe depends on OpenCV, you will need to copy the precompiled OpenCV so
     files into app/src/main/jniLibs. You can download the official OpenCV
     Android SDK from
-    [here](https://github.com/opencv/opencv/releases/download/4.1.0/opencv-4.1.0-android-sdk.zip)
+    [here](https://github.com/opencv/opencv/releases/download/3.4.3/opencv-3.4.3-android-sdk.zip)
     and run:
 
     ```bash
@@ -117,6 +117,7 @@ project.
         implementation 'com.google.code.findbugs:jsr305:3.0.2'
         implementation 'com.google.guava:guava:27.0.1-android'
         implementation 'com.google.guava:guava:27.0.1-android'
+        implementation 'com.google.protobuf:protobuf-lite:3.0.0'
         // CameraX core library
         def camerax_version = "1.0.0-alpha06"
         implementation "androidx.camera:camera-core:$camerax_version"
@@ -125,6 +126,6 @@ project.
     ```
 
 6.  Follow our Android app examples to use MediaPipe in Android Studio for your
-    use case. If you are looking for an example, a working face detection
+    use case. If you are looking for an example, a face detection
     example can be found
-    [here](https://github.com/jiuqiant/mediapipe_aar_example).
+    [here](https://github.com/jiuqiant/mediapipe_face_detection_aar_example) and a multi-hand tracking example can be found [here](https://github.com/jiuqiant/mediapipe_multi_hands_tracking_aar_example).

mediapipe/docs/examples.md

@@ -76,6 +76,15 @@ MediaPipe with a TFLite model for hand tracking in a GPU-accelerated pipeline.
 *   [Android](./hand_tracking_mobile_gpu.md)
 *   [iOS](./hand_tracking_mobile_gpu.md)
 
+### Multi-Hand Tracking with GPU
+
+[Multi-Hand Tracking with GPU](./multi_hand_tracking_mobile_gpu.md) illustrates
+how to use MediaPipe with a TFLite model for multi-hand tracking in a
+GPU-accelerated pipeline.
+
+*   [Android](./multi_hand_tracking_mobile_gpu.md)
+*   [iOS](./multi_hand_tracking_mobile_gpu.md)
+
 ### Hair Segmentation with GPU
 
 [Hair Segmentation on GPU](./hair_segmentation_mobile_gpu.md) illustrates how to
@@ -132,6 +141,15 @@ with live video from a webcam.
 *   [Desktop GPU](./hand_tracking_desktop.md)
 *   [Desktop CPU](./hand_tracking_desktop.md)
 
+### Multi-Hand Tracking on Desktop with Webcam
+
+[Multi-Hand Tracking on Desktop with Webcam](./multi_hand_tracking_desktop.md)
+shows how to use MediaPipe with a TFLite model for multi-hand tracking on
+desktop using CPU or GPU with live video from a webcam.
+
+*   [Desktop GPU](./multi_hand_tracking_desktop.md)
+*   [Desktop CPU](./multi_hand_tracking_desktop.md)
+
 ### Hair Segmentation on Desktop with Webcam
 
 [Hair Segmentation on Desktop with Webcam](./hair_segmentation_desktop.md) shows
@@ -139,3 +157,20 @@ how to use MediaPipe with a TFLite model for hair segmentation on desktop using
 GPU with live video from a webcam.
 
 *   [Desktop GPU](./hair_segmentation_desktop.md)
+
+## Google Coral (machine learning acceleration with Google EdgeTPU)
+
+Below are code samples on how to run MediaPipe on Google Coral Dev Board.
+
+### Object Detection on Coral
+
+[Object Detection on Coral with Webcam](./object_detection_coral_devboard.md)
+shows how to run quantized object detection TFlite model accelerated with
+EdgeTPU on
+[Google Coral Dev Board](https://coral.withgoogle.com/products/dev-board).
+
+### Face Detection on Coral
+
+[Face Detection on Coral with Webcam](./face_detection_coral_devboard.md) shows
+how to use quantized face detection TFlite model accelerated with EdgeTPU on
+[Google Coral Dev Board](https://coral.withgoogle.com/products/dev-board).

mediapipe/docs/face_detection_coral_devboard.md

@@ -0,0 +1,20 @@
+## Face Detection on Coral with Webcam
+
+MediaPipe is able to run cross platform across device types like desktop, mobile
+and edge devices. Here is an example of running MediaPipe
+[face detection pipeline](./face_detection_desktop.md) on edge device like
+[Google Coral dev board](https://coral.withgoogle.com/products/dev-board) with
+[Edge TPU](https://cloud.google.com/edge-tpu/). This MediaPipe Coral face
+detection pipeline is running [coral specific quantized version](https://github.com/google/mediapipe/blob/master/mediapipe/examples/coral/models/face-detector-quantized_edgetpu.tflite)
+of the [MediaPipe face detection TFLite model](https://github.com/google/mediapipe/blob/master/mediapipe/models/face_detection_front.tflite)
+accelerated on Edge TPU.
+
+### Cross compilation of MediaPipe Coral binaries in Docker
+
+We recommend building the MediaPipe binaries not on the edge device due to
+limited compute resulting in long build times. Instead, we will build MediaPipe
+binaries using Docker containers on a more powerful host machine. For step by
+step details of cross compiling and running MediaPipe binaries on Coral dev
+board, please refer to [README.md in MediaPipe Coral example folder](https://github.com/google/mediapipe/blob/master/mediapipe/examples/coral/README.md).
+
+![Face Detection running on Coral](images/face_detection_demo_coral.jpg)

mediapipe/docs/face_detection_desktop.md

@@ -77,7 +77,7 @@ below and paste it into
 ```bash
 # MediaPipe graph that performs face detection with TensorFlow Lite on CPU & GPU.
 # Used in the examples in
-# mediapipie/examples/desktop/face_detection:face_detection_cpu.
+# mediapipe/examples/desktop/face_detection:face_detection_cpu.
 
 # Images on CPU coming into and out of the graph.
 input_stream: "input_video"

mediapipe/docs/gpu.md

@@ -1,10 +1,10 @@
 ## Running on GPUs
 
 -   [Overview](#overview)
--   [OpenGL Support](#graphconfig)
+-   [OpenGL Support](#opengl-support)
 -   [Life of a GPU calculator](#life-of-a-gpu-calculator)
 -   [GpuBuffer to ImageFrame converters](#gpubuffer-to-imageframe-converters)
-
+-   [Disable GPU support](#disable-gpu-support)
 
 ### Overview
 MediaPipe supports calculator nodes for GPU compute and rendering, and allows combining multiple GPU nodes, as well as mixing them with CPU based calculator nodes. There exist several GPU APIs on mobile platforms (eg, OpenGL ES, Metal and Vulkan). MediaPipe does not attempt to offer a single cross-API GPU abstraction. Individual nodes can be written using different APIs, allowing them to take advantage of platform specific features when needed.
@@ -23,6 +23,7 @@ Below are the design principles for GPU support in MediaPipe
    * A calculator should be allowed maximum flexibility in using the GPU for all or part of its operation, combining it with the CPU if necessary.
 
 ### OpenGL support
+
 MediaPipe supports OpenGL ES up to version 3.2 on Android and up to ES 3.0 on iOS. In addition, MediaPipe also supports Metal on iOS.
 
    * MediaPipe allows graphs to run OpenGL in multiple GL contexts. For example, this can be very useful in graphs that combine a slower GPU inference path (eg, at 10 FPS) with a faster GPU rendering path (eg, at 30 FPS): since one GL context corresponds to one sequential command queue, using the same context for both tasks would reduce the rendering frame rate. One challenge MediaPipe's use of multiple contexts solves is the ability to communicate across them. An example scenario is one with an input video that is sent to both the rendering and inferences paths, and rendering needs to have access to the latest output from inference.
@@ -128,3 +129,26 @@ The below diagram shows the data flow in a mobile application that captures vide
 |:--:|
 | *Video frames from the camera are fed into the graph as `GpuBuffer` packets. The input stream is accessed by two calculators in parallel. `GpuBufferToImageFrameCalculator` converts the buffer into an `ImageFrame`, which is then sent through a grayscale converter and a canny filter (both based on OpenCV and running on the CPU), whose output is then converted into a `GpuBuffer` again. A multi-input GPU calculator, GlOverlayCalculator, takes as input both the original `GpuBuffer` and the one coming out of the edge detector, and overlays them using a shader. The output is then sent back to the application using a callback calculator, and the application renders the image to the screen using OpenGL.* |
 
+### Disable GPU Support
+
+By default, building MediaPipe (with no special bazel flags) attempts to compile
+and link against OpenGL/Metal libraries.
+
+There are some command line build flags available to disable/enable GPU support
+within the MediaPipe framework:
+
+```
+# To disable *all* gpu support
+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>
+
+# 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>
+```
+
+Note *MEDIAPIPE_DISABLE_GL_COMPUTE* is automatically defined on all Apple
+systems (Apple doesn't support OpenGL ES 3.1+).
+
+Note on iOS and Android, it is assumed that GPU support will be enabled.

mediapipe/docs/hair_segmentation_desktop.md

@@ -31,8 +31,6 @@ $ bazel build -c opt --copt -DMESA_EGL_NO_X11_HEADERS \
 #INFO: Found 1 target...
 #Target //mediapipe/examples/desktop/hair_segmentation:hair_segmentation_gpu up-to-date:
 #  bazel-bin/mediapipe/examples/desktop/hair_segmentation/hair_segmentation_gpu
-#INFO: Elapsed time: 18.209s, Forge stats: 13026/13057 actions cached, 20.8s CPU used, 0.0s queue time, 89.3 MB ObjFS output (novel bytes: 87.4 MB), 0.0 MB local output, Critical Path: 11.88s, Remote (86.01% of the time): [queue: 0.00%, network: 16.83%, setup: 4.59%, process: 38.92%]
-#INFO: Streaming build results to: http://sponge2/37d5a184-293b-4e98-a43e-b22084db3142
 #INFO: Build completed successfully, 12210 total actions
 
 # This will open up your webcam as long as it is connected and on
@@ -53,7 +51,7 @@ below and paste it into
 ```bash
 # MediaPipe graph that performs hair segmentation with TensorFlow Lite on GPU.
 # Used in the example in
-# mediapipie/examples/android/src/java/com/mediapipe/apps/hairsegmentationgpu.
+# mediapipe/examples/android/src/java/com/mediapipe/apps/hairsegmentationgpu.
 
 # Images on GPU coming into and out of the graph.
 input_stream: "input_video"

mediapipe/docs/hair_segmentation_mobile_gpu.md

@@ -29,7 +29,7 @@ below and paste it into [MediaPipe Visualizer](https://viz.mediapipe.dev/).
 ```bash
 # MediaPipe graph that performs hair segmentation with TensorFlow Lite on GPU.
 # Used in the example in
-# mediapipie/examples/android/src/java/com/mediapipe/apps/hairsegmentationgpu.
+# mediapipe/examples/android/src/java/com/mediapipe/apps/hairsegmentationgpu.
 
 # Images on GPU coming into and out of the graph.
 input_stream: "input_video"

mediapipe/docs/hand_tracking_desktop.md

@@ -31,8 +31,6 @@ $ bazel build -c opt --define MEDIAPIPE_DISABLE_GPU=1 \
 # It should print:
 #Target //mediapipe/examples/desktop/hand_tracking:hand_tracking_cpu up-to-date:
 #  bazel-bin/mediapipe/examples/desktop/hand_tracking/hand_tracking_cpu
-#INFO: Elapsed time: 22.645s, Forge stats: 13356/13463 actions cached, 1.5m CPU used, 0.0s queue time, 819.8 MB ObjFS output (novel bytes: 85.6 MB), 0.0 MB local output, Critical Path: 14.43s, Remote (87.25% of the time): [queue: 0.00%, network: 14.88%, setup: 4.80%, process: 39.80%, fetch: 18.15%]
-#INFO: Streaming build results to: http://sponge2/360196b9-33ab-44b1-84a7-1022b5043307
 #INFO: Build completed successfully, 12517 total actions
 
 # This will open up your webcam as long as it is connected and on
@@ -54,8 +52,6 @@ $ bazel build -c opt --copt -DMESA_EGL_NO_X11_HEADERS \
 # It should print:
 # Target //mediapipe/examples/desktop/hand_tracking:hand_tracking_gpu up-to-date:
 #  bazel-bin/mediapipe/examples/desktop/hand_tracking/hand_tracking_gpu
-#INFO: Elapsed time: 84.055s, Forge stats: 6858/19343 actions cached, 1.6h CPU used, 0.9s queue time, 1.68 GB ObjFS output (novel bytes: 485.1 MB), 0.0 MB local output, Critical Path: 48.14s, Remote (99.40% of the time): [queue: 0.00%, setup: 5.59%, process: 74.44%]
-#INFO: Streaming build results to: http://sponge2/00c7f95f-6fbc-432d-8978-f5d361efca3b
 #INFO: Build completed successfully, 22455 total actions
 
 # This will open up your webcam as long as it is connected and on
@@ -77,7 +73,7 @@ below and paste it into
 # MediaPipe graph that performs hand tracking on desktop with TensorFlow Lite
 # on CPU & GPU.
 # Used in the example in
-# mediapipie/examples/desktop/hand_tracking:hand_tracking_cpu.
+# mediapipe/examples/desktop/hand_tracking:hand_tracking_cpu.
 
 # Images coming into and out of the graph.
 input_stream: "input_video"

mediapipe/docs/hand_tracking_mobile_gpu.md

@@ -100,8 +100,8 @@ see the Visualizing Subgraphs section in the
 ```bash
 # MediaPipe graph that performs hand tracking with TensorFlow Lite on GPU.
 # Used in the examples in
-# mediapipie/examples/android/src/java/com/mediapipe/apps/handtrackinggpu and
+# mediapipe/examples/android/src/java/com/mediapipe/apps/handtrackinggpu and
-# mediapipie/examples/ios/handtrackinggpu.
+# mediapipe/examples/ios/handtrackinggpu.
 
 # Images coming into and out of the graph.
 input_stream: "input_video"

mediapipe/docs/hello_world_android.md

@@ -629,7 +629,7 @@ to load both dependencies:
 static {
   // Load all native libraries needed by the app.
   System.loadLibrary("mediapipe_jni");
-  System.loadLibrary("opencv_java4");
+  System.loadLibrary("opencv_java3");
 }
 ```
 

mediapipe/docs/install.md

@@ -7,11 +7,8 @@ future.
 Note: If you plan to use TensorFlow calculators and example apps, there is a
 known issue with gcc and g++ version 6.3 and 7.3. Please use other versions.
 
-Note: While Mediapipe configures TensorFlow, if you see the
+Note: To make Mediapipe work with TensorFlow, please install the python "future"
-following error:
+library and the python "six" library using `pip install --user future six`.
-`"...git_configure.bzl", line 14, in _fail fail(("%sGit Configuration
-Error:%s %...)))`,
-please install the python future library using: `$ pip install --user future`.
 
 Choose your operating system:
 
@@ -42,11 +39,19 @@ To build and run iOS apps:
     $ cd mediapipe
     ```
 
-2.  Install Bazel (version between 0.24.1 and 0.29.1).
+2.  Install Bazel (0.24.1 and above required).
 
-    Follow the official
+    Option 1. Use package manager tool to install the latest version of Bazel.
-    [documentation](https://docs.bazel.build/versions/master/install-ubuntu.html)
+
-    to install Bazel manually. Note that MediaPipe doesn't support Bazel 1.0.0+ yet.
+    ```bash
+    $ sudo apt-get install bazel
+
+    # Run 'bazel version' to check version of bazel installed
+    ```
+
+    Option 2. Follow the official
+    [Bazel documentation](https://docs.bazel.build/versions/master/install-ubuntu.html)
+    to install any version of Bazel manually.
 
 3.  Install OpenCV and FFmpeg.
 
@@ -152,11 +157,11 @@ To build and run iOS apps:
     $ cd mediapipe
     ```
 
-2.  Install Bazel (version between 0.24.1 and 0.29.1).
+2.  Install Bazel (0.24.1 and above required).
 
     Follow the official
-    [documentation](https://docs.bazel.build/versions/master/install-redhat.html)
+    [Bazel documentation](https://docs.bazel.build/versions/master/install-redhat.html)
-    to install Bazel manually. Note that MediaPipe doesn't support Bazel 1.0.0+ yet.
+    to install Bazel manually.
 
 3.  Install OpenCV.
 
@@ -240,24 +245,23 @@ To build and run iOS apps:
     $ cd mediapipe
     ```
 
-3.  Install Bazel (version between 0.24.1 and 0.29.1).
+3.  Install Bazel (version between 0.24.1 and 1.1.0).
 
-     Option 1. Use package manager tool to install Bazel 0.29.1
+    Option 1. Use package manager tool to install Bazel 1.1.0
 
     ```bash
-    # If Bazel 1.0.0+ was installed.
+    # If Bazel 1.1.0+ was installed.
     $ brew uninstall bazel
-
+    # Install Bazel 1.1.0
-    # Install Bazel 0.29.1
+    $ brew install https://raw.githubusercontent.com/bazelbuild/homebrew-tap/f8a0fa981bcb1784a0d0823e14867b844e94fb3d/Formula/bazel.rb
-    $ brew install https://raw.githubusercontent.com/bazelbuild/homebrew-tap/223ffb570c21c0a2af251afc6df9dec0214c6e74/Formula/bazel.rb
     $ brew link bazel
-
+    # Run 'bazel version' to check version of bazel
-    # Run 'bazel version' to check version of bazel installed
     ```
 
     Option 2. Follow the official
-    [documentation](https://docs.bazel.build/versions/master/install-os-x.html#install-with-installer-mac-os-x)
+    [Bazel documentation](https://docs.bazel.build/versions/master/install-os-x.html#install-with-installer-mac-os-x)
-    to install Bazel manually. Note that MediaPipe doesn't support Bazel 1.0.0+ yet.
+    to install any version of Bazel manually. Note that MediaPipe doesn't
+    support Bazel 1.1.0+ on macOS yet.
 
 4.  Install OpenCV and FFmpeg.
 
@@ -526,7 +530,7 @@ This will use a Docker image that will isolate mediapipe's installation from the
     ```bash
     $ docker run -it --name mediapipe mediapipe:latest
 
-    root@bca08b91ff63:/mediapipe# bash ./setup_android_sdk_and_ndk
+    root@bca08b91ff63:/mediapipe# bash ./setup_android_sdk_and_ndk.sh
 
     # Should print:
     # Android NDK is now installed. Consider setting $ANDROID_NDK_HOME environment variable to be /root/Android/Sdk/ndk-bundle/android-ndk-r18b
@@ -579,6 +583,11 @@ export ANDROID_HOME=<path to the Android SDK>
 export ANDROID_NDK_HOME=<path to the Android NDK>
 ```
 
+In order to use MediaPipe on earlier Android versions, MediaPipe needs to switch
+to a lower Android API level. You can achieve this by specifying `api_level =
+<api level integer>` in android_ndk_repository() and/or android_sdk_repository()
+in the [`WORKSPACE`] file.
+
 Please verify all the necessary packages are installed.
 
 *   Android SDK Platform API Level 28 or 29

mediapipe/docs/mediapipe_ios_setup.md

@@ -7,7 +7,8 @@
 
 2.  Install [Bazel](https://bazel.build/).
 
-    See their [instructions](https://docs.bazel.build/versions/master/install-os-x.html).
+    See their
+    [instructions](https://docs.bazel.build/versions/master/install-os-x.html).
     We recommend using [Homebrew](https://brew.sh/):
 
     ```bash
@@ -15,13 +16,23 @@
     brew install bazelbuild/tap/bazel
     ```
 
-3.  Clone the MediaPipe repository.
+3.  Install python "future" and "six".
+
+    To make Mediapipe work with TensorFlow, please install the python "future"
+    library and the python "six" library:
+
+    ```bash
+    pip install --user future six
+    ```
+
+4.  Clone the MediaPipe repository.
 
     ```bash
     git clone https://github.com/google/mediapipe.git
     ```
 
-4.  Symlink or copy your provisioning profile to `mediapipe/mediapipe/provisioning_profile.mobileprovision`.
+5.  Symlink or copy your provisioning profile to
+    `mediapipe/mediapipe/provisioning_profile.mobileprovision`.
 
     ```bash
     cd mediapipe

mediapipe/docs/multi_hand_tracking_desktop.md

@@ -0,0 +1,177 @@
+## Multi-Hand Tracking on Desktop
+
+This is an example of using MediaPipe to run hand tracking models (TensorFlow
+Lite) and render bounding boxes on the detected hand instances (for multiple
+hands). To know more about the hand tracking models, please refer to the model
+[`README file`]. Moreover, if you are interested in running the same TensorfFlow
+Lite model on Android/iOS, please see the
+[Mulit-Hand Tracking on GPU on Android/iOS](multi_hand_tracking_mobile_gpu.md)
+and
+
+We show the hand tracking demos with TensorFlow Lite model using the Webcam:
+
+-   [TensorFlow Lite Multi-Hand Tracking Demo with Webcam (CPU)](#tensorflow-lite-multi-hand-tracking-demo-with-webcam-cpu)
+
+-   [TensorFlow Lite Multi-Hand Tracking Demo with Webcam (GPU)](#tensorflow-lite-multi-hand-tracking-demo-with-webcam-gpu)
+
+Note: Desktop GPU works only on Linux. Mesa drivers need to be installed. Please
+see
+[step 4 of "Installing on Debian and Ubuntu" in the installation guide](./install.md).
+
+Note: If MediaPipe depends on OpenCV 2, please see the
+[known issues with OpenCV 2](#known-issues-with-opencv-2) section.
+
+### TensorFlow Lite Multi-Hand Tracking Demo with Webcam (CPU)
+
+To build and run the TensorFlow Lite example on desktop (CPU) with Webcam, run:
+
+```bash
+# Video from webcam running on desktop CPU
+$ bazel build -c opt --define MEDIAPIPE_DISABLE_GPU=1 \
+    mediapipe/examples/desktop/multi_hand_tracking:multi_hand_tracking_cpu
+
+# It should print:
+#Target //mediapipe/examples/desktop/multi_hand_tracking:multi_hand_tracking_cpu up-to-date:
+#  bazel-bin/mediapipe/examples/desktop/multi_hand_tracking/multi_hand_tracking_cpu
+
+# This will open up your webcam as long as it is connected and on
+# Any errors is likely due to your webcam being not accessible
+$ GLOG_logtostderr=1 bazel-bin/mediapipe/examples/desktop/multi_hand_tracking/multi_hand_tracking_cpu \
+    --calculator_graph_config_file=mediapipe/graphs/hand_tracking/multi_hand_tracking_desktop_live.pbtxt
+```
+
+### TensorFlow Lite Multi-Hand Tracking Demo with Webcam (GPU)
+
+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 \
+    mediapipe/examples/desktop/multi_hand_tracking:multi_hand_tracking_gpu
+
+# It should print:
+# Target //mediapipe/examples/desktop/multi_hand_tracking:multi_hand_tracking_gpu up-to-date:
+#  bazel-bin/mediapipe/examples/desktop/multi_hand_tracking/multi_hand_tracking_gpu
+
+# This will open up your webcam as long as it is connected and on
+# Any errors is likely due to your webcam being not accessible,
+# or GPU drivers not setup properly.
+$ GLOG_logtostderr=1 bazel-bin/mediapipe/examples/desktop/multi_hand_tracking/multi_hand_tracking_gpu \
+    --calculator_graph_config_file=mediapipe/graphs/hand_tracking/multi_hand_tracking_mobile.pbtxt
+```
+
+#### Graph
+
+![graph visualization](images/multi_hand_tracking_desktop.png)
+
+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).
+
+```bash
+# MediaPipe graph that performs multi-hand tracking on desktop with TensorFlow
+# Lite on CPU.
+# Used in the example in
+# mediapipie/examples/desktop/multi_hand_tracking:multi_hand_tracking_cpu.
+
+# Images coming into and out of the graph.
+input_stream: "input_video"
+output_stream: "output_video"
+
+# Determines if an input vector of NormalizedRect has a size greater than or
+# equal to the provided min_size.
+node {
+  calculator: "NormalizedRectVectorHasMinSizeCalculator"
+  input_stream: "ITERABLE:prev_multi_hand_rects_from_landmarks"
+  output_stream: "prev_has_enough_hands"
+  node_options: {
+    [type.googleapis.com/mediapipe.CollectionHasMinSizeCalculatorOptions] {
+      # This value can be changed to support tracking arbitrary number of hands.
+      # Please also remember to modify max_vec_size in
+      # ClipVectorSizeCalculatorOptions in
+      # mediapipe/graphs/hand_tracking/subgraphs/multi_hand_detection_gpu.pbtxt
+      min_size: 2
+    }
+  }
+}
+
+# Drops the incoming image if the previous frame had at least N hands.
+# Otherwise, passes the incoming image through to trigger a new round of hand
+# detection in MultiHandDetectionSubgraph.
+node {
+  calculator: "GateCalculator"
+  input_stream: "input_video"
+  input_stream: "DISALLOW:prev_has_enough_hands"
+  output_stream: "multi_hand_detection_input_video"
+  node_options: {
+    [type.googleapis.com/mediapipe.GateCalculatorOptions] {
+      empty_packets_as_allow: true
+    }
+  }
+}
+
+# Subgraph that detections hands (see multi_hand_detection_cpu.pbtxt).
+node {
+  calculator: "MultiHandDetectionSubgraph"
+  input_stream: "multi_hand_detection_input_video"
+  output_stream: "DETECTIONS:multi_palm_detections"
+  output_stream: "NORM_RECTS:multi_palm_rects"
+}
+
+# Subgraph that localizes hand landmarks for multiple hands (see
+# multi_hand_landmark.pbtxt).
+node {
+  calculator: "MultiHandLandmarkSubgraph"
+  input_stream: "IMAGE:input_video"
+  input_stream: "NORM_RECTS:multi_hand_rects"
+  output_stream: "LANDMARKS:multi_hand_landmarks"
+  output_stream: "NORM_RECTS:multi_hand_rects_from_landmarks"
+}
+
+# Caches a hand rectangle fed back from MultiHandLandmarkSubgraph, and upon the
+# arrival of the next input image sends out the cached rectangle with the
+# timestamp replaced by that of the input image, essentially generating a packet
+# that carries the previous hand rectangle. Note that upon the arrival of the
+# very first input image, an empty packet is sent out to jump start the
+# feedback loop.
+node {
+  calculator: "PreviousLoopbackCalculator"
+  input_stream: "MAIN:input_video"
+  input_stream: "LOOP:multi_hand_rects_from_landmarks"
+  input_stream_info: {
+    tag_index: "LOOP"
+    back_edge: true
+  }
+  output_stream: "PREV_LOOP:prev_multi_hand_rects_from_landmarks"
+}
+
+# Performs association between NormalizedRect vector elements from previous
+# frame and those from the current frame if MultiHandDetectionSubgraph runs.
+# This calculator ensures that the output multi_hand_rects vector doesn't
+# contain overlapping regions based on the specified min_similarity_threshold.
+node {
+  calculator: "AssociationNormRectCalculator"
+  input_stream: "prev_multi_hand_rects_from_landmarks"
+  input_stream: "multi_palm_rects"
+  output_stream: "multi_hand_rects"
+  node_options: {
+    [type.googleapis.com/mediapipe.AssociationCalculatorOptions] {
+      min_similarity_threshold: 0.5
+    }
+  }
+}
+
+# Subgraph that renders annotations and overlays them on top of the input
+# images (see multi_hand_renderer_cpu.pbtxt).
+node {
+  calculator: "MultiHandRendererSubgraph"
+  input_stream: "IMAGE:input_video"
+  input_stream: "DETECTIONS:multi_palm_detections"
+  input_stream: "LANDMARKS:multi_hand_landmarks"
+  input_stream: "NORM_RECTS:0:multi_palm_rects"
+  input_stream: "NORM_RECTS:1:multi_hand_rects"
+  output_stream: "IMAGE:output_video"
+}
+```
+
+[`README file`]:https://github.com/google/mediapipe/tree/master/mediapipe/README.md

mediapipe/docs/multi_hand_tracking_mobile_gpu.md

@@ -0,0 +1,755 @@
+# Multi-Hand Tracking (GPU)
+
+This doc focuses on the
+[example graph](https://github.com/google/mediapipe/tree/master/mediapipe/graphs/hand_tracking/multi_hand_tracking_mobile.pbtxt)
+that performs multi-hand tracking with TensorFlow Lite on GPU. It is related to
+the [hand_tracking_example](./hand_tracking_mobile_gpu.md), and we recommend
+users to review the (single) hand tracking example first.
+
+![multi_hand_tracking_android_gpu.gif](images/mobile/multi_hand_tracking_android_gpu.gif)
+
+In the visualization above, the red dots represent the hand landmarks and the
+green lines are simply connections between selected landmark paris for
+visualization of the hand skeleton. When there are fewer than `N` hands (`N=2`
+in the graphs here), the purple box represents a hand rectangle that covers the
+entire hand, derived from hand detection (see
+[hand_detection_example](./hand_detection_mobile_gpu.md)). When there are `N`
+hands (i.e. 2 hands for the graphs here), the red boxes represent hand
+rectangles for each of the hands, derived from the previous round of hand
+landmark localization using an ML model (see also
+[model card](https://mediapipe.page.link/handmc)). Hand landmark localization
+for each hand is performed only within the hand rectangle for computational
+efficiency and accuracy. Hand detection is only invoked whenever there are fewer
+than `N` hands in the previous iteration.
+
+This example can also run a model that localizes hand landmarks in 3D (i.e.,
+estimating an extra z coordinate):
+
+![multi_hand_tracking_3d_android_gpu.gif](images/mobile/multi_hand_tracking_3d_android_gpu.gif)
+
+In the visualization above, the localized hand landmarks are represented by dots
+in different shades, with the brighter ones denoting landmarks closer to the
+camera.
+
+## Android
+
+[Source](https://github.com/google/mediapipe/tree/master/mediapipe/examples/android/src/java/com/google/mediapipe/apps/multihandtrackinggpu)
+
+To build the app yourself, run:
+
+```bash
+bazel build -c opt --config=android_arm64 mediapipe/examples/android/src/java/com/google/mediapipe/apps/multihandtrackinggpu
+```
+
+To build for the 3D mode, run:
+
+```bash
+bazel build -c opt --config=android_arm64 --define 3D=true mediapipe/examples/android/src/java/com/google/mediapipe/apps/multihandtrackinggpu
+```
+
+Once the app is built, install it on Android device with:
+
+```bash
+adb install bazel-bin/mediapipe/examples/android/src/java/com/google/mediapipe/apps/multihandtrackinggpu/multihandtrackinggpu.apk
+```
+
+## iOS
+
+[Source](https://github.com/google/mediapipe/tree/master/mediapipe/examples/ios/multihandtrackinggpu).
+
+See the general [instructions](./mediapipe_ios_setup.md) for building iOS
+examples and generating an Xcode project. This will be the HandDetectionGpuApp
+target.
+
+To build on the command line:
+
+```bash
+bazel build -c opt --config=ios_arm64 mediapipe/examples/ios/multihandtrackinggpu:MultiHandTrackingGpuApp
+```
+
+To build for the 3D mode, run:
+
+```bash
+bazel build -c opt --config=ios_arm64 --define 3D=true mediapipe/examples/ios/multihandtrackinggpu:MultiHandTrackingGpuApp
+```
+
+## Graph
+
+The multi-hand tracking [main graph](#main-graph) internal utilizes a
+[multi_hand_detection_subgraph](#multi-hand-detection-subgraph), a
+[multi_hand_landmark_subgraph](#multi-hand-landmark-subgraph), and a
+[multi_hand_renderer_subgraph](#multi-hand-renderer-subgraph).
+
+The subgraphs show up in the main graph visualization as nodes colored in
+purple, and the subgraph itself can also be visualized just like a regular
+graph. For more information on how to visualize a graph that includes subgraphs,
+see the Visualizing Subgraphs section in the
+[visualizer documentation](./visualizer.md).
+
+### Main Graph
+
+![multi_hand_tracking_mobile_graph](images/mobile/multi_hand_tracking_mobile.png)
+
+There are two key differences between this graph and the
+[single_hand_tracking_mobile_graph](./hand_tracking_mobile_gpu.md).
+
+1.  There is a `NormalizedRectVectorHasMinSize` calculator, that checks if in
+    input vector of `NormalizedRect` objects has a minimum size equal to `N`. In
+    this graph, if the vector contains fewer than `N` objects,
+    `MultiHandDetection` subgraph runs. Otherwise, the `GateCalculator` doesn't
+    send any image packets to the `MultiHandDetection` subgraph. This way, the
+    main graph is efficient in that it avoids running the costly hand detection
+    step when there are already `N` hands in the frame.
+2.  The `MergeCalculator` has been replaced by the `AssociationNormRect`
+    calculator. This `AssociationNormRect` takes as input a vector of
+    `NormalizedRect` objects from the `MultiHandDetection` subgraph on the
+    current frame, and a vector of `NormalizedRect` objects from the
+    `MultiHandLandmark` subgraph from the previous frame, and performs an
+    association operation between these objects. This calculator ensures that
+    the output vector doesn't contain overlapping regions based on the specified
+    `min_similarity_threshold`.
+
+[Source pbtxt file](https://github.com/google/mediapipe/tree/master/mediapipe/graphs/hand_tracking/multi_hand_tracking_mobile.pbtxt)
+
+```bash
+# MediaPipe graph that performs multi-hand tracking with TensorFlow Lite on GPU.
+# Used in the examples in
+# mediapipie/examples/android/src/java/com/mediapipe/apps/multihandtrackinggpu.
+
+# Images coming into and out of the graph.
+input_stream: "input_video"
+output_stream: "output_video"
+
+# Throttles the images flowing downstream for flow control. It passes through
+# the very first incoming image unaltered, and waits for downstream nodes
+# (calculators and subgraphs) in the graph to finish their tasks before it
+# passes through another image. All images that come in while waiting are
+# dropped, limiting the number of in-flight images in most part of the graph to
+# 1. This prevents the downstream nodes from queuing up incoming images and data
+# excessively, which leads to increased latency and memory usage, unwanted in
+# real-time mobile applications. It also eliminates unnecessarily computation,
+# e.g., the output produced by a node may get dropped downstream if the
+# subsequent nodes are still busy processing previous inputs.
+node {
+  calculator: "FlowLimiterCalculator"
+  input_stream: "input_video"
+  input_stream: "FINISHED:multi_hand_rects"
+  input_stream_info: {
+    tag_index: "FINISHED"
+    back_edge: true
+  }
+  output_stream: "throttled_input_video"
+}
+
+# Determines if an input vector of NormalizedRect has a size greater than or
+# equal to the provided min_size.
+node {
+  calculator: "NormalizedRectVectorHasMinSizeCalculator"
+  input_stream: "ITERABLE:prev_multi_hand_rects_from_landmarks"
+  output_stream: "prev_has_enough_hands"
+  node_options: {
+    [type.googleapis.com/mediapipe.CollectionHasMinSizeCalculatorOptions] {
+      # This value can be changed to support tracking arbitrary number of hands.
+      # Please also remember to modify max_vec_size in
+      # ClipVectorSizeCalculatorOptions in
+      # mediapipe/graphs/hand_tracking/subgraphs/multi_hand_detection_gpu.pbtxt
+      min_size: 2
+    }
+  }
+}
+
+# Drops the incoming image if the previous frame had at least N hands.
+# Otherwise, passes the incoming image through to trigger a new round of hand
+# detection in MultiHandDetectionSubgraph.
+node {
+  calculator: "GateCalculator"
+  input_stream: "throttled_input_video"
+  input_stream: "DISALLOW:prev_has_enough_hands"
+  output_stream: "multi_hand_detection_input_video"
+  node_options: {
+    [type.googleapis.com/mediapipe.GateCalculatorOptions] {
+      empty_packets_as_allow: true
+    }
+  }
+}
+
+# Subgraph that detections hands (see multi_hand_detection_gpu.pbtxt).
+node {
+  calculator: "MultiHandDetectionSubgraph"
+  input_stream: "multi_hand_detection_input_video"
+  output_stream: "DETECTIONS:multi_palm_detections"
+  output_stream: "NORM_RECTS:multi_palm_rects"
+}
+
+# Subgraph that localizes hand landmarks for multiple hands (see
+# multi_hand_landmark.pbtxt).
+node {
+  calculator: "MultiHandLandmarkSubgraph"
+  input_stream: "IMAGE:throttled_input_video"
+  input_stream: "NORM_RECTS:multi_hand_rects"
+  output_stream: "LANDMARKS:multi_hand_landmarks"
+  output_stream: "NORM_RECTS:multi_hand_rects_from_landmarks"
+}
+
+# Caches a hand rectangle fed back from MultiHandLandmarkSubgraph, and upon the
+# arrival of the next input image sends out the cached rectangle with the
+# timestamp replaced by that of the input image, essentially generating a packet
+# that carries the previous hand rectangle. Note that upon the arrival of the
+# very first input image, an empty packet is sent out to jump start the
+# feedback loop.
+node {
+  calculator: "PreviousLoopbackCalculator"
+  input_stream: "MAIN:throttled_input_video"
+  input_stream: "LOOP:multi_hand_rects_from_landmarks"
+  input_stream_info: {
+    tag_index: "LOOP"
+    back_edge: true
+  }
+  output_stream: "PREV_LOOP:prev_multi_hand_rects_from_landmarks"
+}
+
+# Performs association between NormalizedRect vector elements from previous
+# frame and those from the current frame if MultiHandDetectionSubgraph runs.
+# This calculator ensures that the output multi_hand_rects vector doesn't
+# contain overlapping regions based on the specified min_similarity_threshold.
+node {
+  calculator: "AssociationNormRectCalculator"
+  input_stream: "prev_multi_hand_rects_from_landmarks"
+  input_stream: "multi_palm_rects"
+  output_stream: "multi_hand_rects"
+  node_options: {
+    [type.googleapis.com/mediapipe.AssociationCalculatorOptions] {
+      min_similarity_threshold: 0.5
+    }
+  }
+}
+
+# Subgraph that renders annotations and overlays them on top of the input
+# images (see multi_hand_renderer_gpu.pbtxt).
+node {
+  calculator: "MultiHandRendererSubgraph"
+  input_stream: "IMAGE:throttled_input_video"
+  input_stream: "DETECTIONS:multi_palm_detections"
+  input_stream: "LANDMARKS:multi_hand_landmarks"
+  input_stream: "NORM_RECTS:0:multi_palm_rects"
+  input_stream: "NORM_RECTS:1:multi_hand_rects"
+  output_stream: "IMAGE:output_video"
+}
+```
+
+### Multi-Hand Detection Subgraph
+
+![multi_hand_detection_gpu_subgraph](images/mobile/multi_hand_detection_gpu_subgraph.png)
+
+This graph outputs a vector of `NormalizedRect` objects corresponding to each of
+the hand instances visible in the frame. Note that at the end of this graph,
+there is a `ClipNormalizedRectVectorSizeCalculator`. This calculator clips the
+size of the input vector to a maximum size `N`. This implies that the
+`MultiHandDetection` subgraph outputs a vector of maximum `N` hand instance
+locations.
+
+[Source pbtxt file](https://github.com/google/mediapipe/tree/master/mediapipe/graphs/hand_tracking/subgraphs/multi_hand_detection_gpu.pbtxt)
+
+```bash
+# MediaPipe multi-hand detection subgraph.
+
+type: "MultiHandDetectionSubgraph"
+
+input_stream: "input_video"
+output_stream: "DETECTIONS:palm_detections"
+output_stream: "NORM_RECTS:clipped_hand_rects_from_palm_detections"
+
+# Transforms the input image on GPU to a 256x256 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"
+  output_stream: "LETTERBOX_PADDING:letterbox_padding"
+  node_options: {
+    [type.googleapis.com/mediapipe.ImageTransformationCalculatorOptions] {
+      output_width: 256
+      output_height: 256
+      scale_mode: FIT
+    }
+  }
+}
+
+# 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
+    }
+  }
+}
+
+# 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"
+}
+
+# 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_GPU:detection_tensors"
+  input_side_packet: "CUSTOM_OP_RESOLVER:opresolver"
+  node_options: {
+    [type.googleapis.com/mediapipe.TfLiteInferenceCalculatorOptions] {
+      model_path: "mediapipe/models/palm_detection.tflite"
+      use_gpu: true
+    }
+  }
+}
+
+# Generates a single side packet containing a vector of SSD anchors based on
+# the specification in the options.
+node {
+  calculator: "SsdAnchorsCalculator"
+  output_side_packet: "anchors"
+  node_options: {
+    [type.googleapis.com/mediapipe.SsdAnchorsCalculatorOptions] {
+      num_layers: 5
+      min_scale: 0.1171875
+      max_scale: 0.75
+      input_size_height: 256
+      input_size_width: 256
+      anchor_offset_x: 0.5
+      anchor_offset_y: 0.5
+      strides: 8
+      strides: 16
+      strides: 32
+      strides: 32
+      strides: 32
+      aspect_ratios: 1.0
+      fixed_anchor_size: true
+    }
+  }
+}
+
+# Decodes the detection tensors generated by the TensorFlow Lite model, based on
+# the SSD anchors and the specification in the options, into a vector of
+# detections. Each detection describes a detected object.
+node {
+  calculator: "TfLiteTensorsToDetectionsCalculator"
+  input_stream: "TENSORS_GPU:detection_tensors"
+  input_side_packet: "ANCHORS:anchors"
+  output_stream: "DETECTIONS:detections"
+  node_options: {
+    [type.googleapis.com/mediapipe.TfLiteTensorsToDetectionsCalculatorOptions] {
+      num_classes: 1
+      num_boxes: 2944
+      num_coords: 18
+      box_coord_offset: 0
+      keypoint_coord_offset: 4
+      num_keypoints: 7
+      num_values_per_keypoint: 2
+      sigmoid_score: true
+      score_clipping_thresh: 100.0
+      reverse_output_order: true
+
+      x_scale: 256.0
+      y_scale: 256.0
+      h_scale: 256.0
+      w_scale: 256.0
+      min_score_thresh: 0.7
+    }
+  }
+}
+
+# Performs non-max suppression to remove excessive detections.
+node {
+  calculator: "NonMaxSuppressionCalculator"
+  input_stream: "detections"
+  output_stream: "filtered_detections"
+  node_options: {
+    [type.googleapis.com/mediapipe.NonMaxSuppressionCalculatorOptions] {
+      min_suppression_threshold: 0.3
+      overlap_type: INTERSECTION_OVER_UNION
+      algorithm: WEIGHTED
+      return_empty_detections: true
+    }
+  }
+}
+
+# Maps detection label IDs to the corresponding label text ("Palm"). The label
+# map is provided in the label_map_path option.
+node {
+  calculator: "DetectionLabelIdToTextCalculator"
+  input_stream: "filtered_detections"
+  output_stream: "labeled_detections"
+  node_options: {
+    [type.googleapis.com/mediapipe.DetectionLabelIdToTextCalculatorOptions] {
+      label_map_path: "mediapipe/models/palm_detection_labelmap.txt"
+    }
+  }
+}
+
+# Adjusts detection locations (already normalized to [0.f, 1.f]) on the
+# letterboxed image (after image transformation with the FIT scale mode) to the
+# corresponding locations on the same image with the letterbox removed (the
+# input image to the graph before image transformation).
+node {
+  calculator: "DetectionLetterboxRemovalCalculator"
+  input_stream: "DETECTIONS:labeled_detections"
+  input_stream: "LETTERBOX_PADDING:letterbox_padding"
+  output_stream: "DETECTIONS:palm_detections"
+}
+
+# Extracts image size from the input images.
+node {
+  calculator: "ImagePropertiesCalculator"
+  input_stream: "IMAGE_GPU:input_video"
+  output_stream: "SIZE:image_size"
+}
+
+# Converts each palm detection into a rectangle (normalized by image size)
+# that encloses the palm and is rotated such that the line connecting center of
+# the wrist and MCP of the middle finger is aligned with the Y-axis of the
+# rectangle.
+node {
+  calculator: "DetectionsToRectsCalculator"
+  input_stream: "DETECTIONS:palm_detections"
+  input_stream: "IMAGE_SIZE:image_size"
+  output_stream: "NORM_RECTS:palm_rects"
+  node_options: {
+    [type.googleapis.com/mediapipe.DetectionsToRectsCalculatorOptions] {
+      rotation_vector_start_keypoint_index: 0  # Center of wrist.
+      rotation_vector_end_keypoint_index: 2  # MCP of middle finger.
+      rotation_vector_target_angle_degrees: 90
+      output_zero_rect_for_empty_detections: true
+    }
+  }
+}
+
+# Expands and shifts the rectangle that contains the palm so that it's likely
+# to cover the entire hand.
+node {
+  calculator: "RectTransformationCalculator"
+  input_stream: "NORM_RECTS:palm_rects"
+  input_stream: "IMAGE_SIZE:image_size"
+  output_stream: "hand_rects_from_palm_detections"
+  node_options: {
+    [type.googleapis.com/mediapipe.RectTransformationCalculatorOptions] {
+      scale_x: 2.6
+      scale_y: 2.6
+      shift_y: -0.5
+      square_long: true
+    }
+  }
+}
+
+# Clips the size of the input vector to the provided max_vec_size. This
+# determines the maximum number of hand instances this graph outputs.
+# Note that the performance gain of clipping detections earlier in this graph is
+# minimal because NMS will minimize overlapping detections and the number of
+# detections isn't expected to exceed 5-10.
+node {
+  calculator: "ClipNormalizedRectVectorSizeCalculator"
+  input_stream: "hand_rects_from_palm_detections"
+  output_stream: "clipped_hand_rects_from_palm_detections"
+  node_options: {
+    [type.googleapis.com/mediapipe.ClipVectorSizeCalculatorOptions] {
+      # This value can be changed to support tracking arbitrary number of hands.
+      # Please also remember to modify min_size in
+      # CollectionHsMinSizeCalculatorOptions in
+      # mediapipe/graphs/hand_tracking/multi_hand_tracking_mobile.pbtxt and
+      # mediapipe/graphs/hand_tracking/multi_hand_tracking_desktop_live.pbtxt.
+      max_vec_size: 2
+    }
+  }
+}
+```
+
+### Multi-Hand Landmark Subgraph
+
+![multi_hand_landmark_subgraph.pbtxt](images/mobile/multi_hand_landmark_subgraph.png)
+
+This graph accepts as input a vector of `NormalizedRect` objects, corresponding
+the the region of each hand instance in the input image. For each
+`NormalizedRect` object, the graph runs the existing `HandLandmark` subgraph and
+collect the outputs of this subgraph into vectors. This is enabled by
+`BeginLoop` and `EndLoop` calculators.
+
+The `BeginLoop` calculator accepts as input a packet containing an iterable
+collection of elements. This calculator is templatized (see
+[begin_loop_calculator.h](https://github.com/google/mediapipe/tree/master/mediapipe/calculators/core/begin_loop_calculator.h)).
+If the input packet arrived at a timestamp `ts`, this calculator outputs each
+element in the collection at a fake timestamp `internal_ts`. At the end of the
+collection, the calculator outputs the arrival timestamp `ts` in the output
+stream tagged with `BATCH_END`.
+
+The nodes between the `BeginLoop` calculator and the corresponding `EndLoop`
+calculator process individual packets at the fake timestamps `internal_ts`.
+After each element is processed, it is sent to the `EndLoop` calculator (see
+[end_loop_calculator.h](https://github.com/google/mediapipe/tree/master/mediapipe/calculators/core/end_loop_calculator.h)),
+which collects these elements in an output collection. The `EndLoop` calculator
+listens for packets from the `BATCH_END` output stream of the `BeginLoop`
+calculator. When the `BATCH_END` packet containing the real timestamp `ts`
+arrives at the `EndLoop` calculator, the `EndLoop` calculator outputs a packet
+containing the collection of processed elements at the real timestamp `ts`.
+
+In the multi-hand landmark subgraph, the `EndLoop` calculators collect the
+output vector of hand landmarks per hand instance, the boolean values indicating
+the presence of each hand and the `NormalizedRect` objects corresponding to the
+regions surrounding each hand into vectors.
+
+Finally, based on the hand presence boolean value, the graph filters the
+collections of hand landmarks and `NormalizdRect` objects corresponding to each
+hand instance.
+
+[Source pbtxt file](https://github.com/google/mediapipe/tree/master/mediapipe/graphs/hand_tracking/subgraphs/multi_hand_landmark.pbtxt)
+
+```bash
+# MediaPipe hand landmark localization subgraph.
+
+type: "MultiHandLandmarkSubgraph"
+
+input_stream: "IMAGE:input_video"
+# A vector of NormalizedRect, one per each hand detected.
+input_stream: "NORM_RECTS:multi_hand_rects"
+# A vector of NormalizedLandmarks, one set per each hand.
+output_stream: "LANDMARKS:filtered_multi_hand_landmarks"
+# A vector of NormalizedRect, one per each hand.
+output_stream: "NORM_RECTS:filtered_multi_hand_rects_for_next_frame"
+
+# Outputs each element of multi_hand_rects at a fake timestamp for the rest
+# of the graph to process. Clones the input_video packet for each
+# single_hand_rect at the fake timestamp. At the end of the loop,
+# outputs the BATCH_END timestamp for downstream calculators to inform them
+# that all elements in the vector have been processed.
+node {
+  calculator: "BeginLoopNormalizedRectCalculator"
+  input_stream: "ITERABLE:multi_hand_rects"
+  input_stream: "CLONE:input_video"
+  output_stream: "ITEM:single_hand_rect"
+  output_stream: "CLONE:input_video_cloned"
+  output_stream: "BATCH_END:single_hand_rect_timestamp"
+}
+
+node {
+  calculator: "HandLandmarkSubgraph"
+  input_stream: "IMAGE:input_video_cloned"
+  input_stream: "NORM_RECT:single_hand_rect"
+  output_stream: "LANDMARKS:single_hand_landmarks"
+  output_stream: "NORM_RECT:single_hand_rect_from_landmarks"
+  output_stream: "PRESENCE:single_hand_presence"
+}
+
+# Collects the boolean presence value for each single hand into a vector. Upon
+# receiving the BATCH_END timestamp, outputs a vector of boolean values at the
+# BATCH_END timestamp.
+node {
+  calculator: "EndLoopBooleanCalculator"
+  input_stream: "ITEM:single_hand_presence"
+  input_stream: "BATCH_END:single_hand_rect_timestamp"
+  output_stream: "ITERABLE:multi_hand_presence"
+}
+
+# Collects a set of landmarks for each hand into a vector. Upon receiving the
+# BATCH_END timestamp, outputs the vector of landmarks at the BATCH_END
+# timestamp.
+node {
+  calculator: "EndLoopNormalizedLandmarkListVectorCalculator"
+  input_stream: "ITEM:single_hand_landmarks"
+  input_stream: "BATCH_END:single_hand_rect_timestamp"
+  output_stream: "ITERABLE:multi_hand_landmarks"
+}
+
+# Collects a NormalizedRect for each hand into a vector. Upon receiving the
+# BATCH_END timestamp, outputs the vector of NormalizedRect at the BATCH_END
+# timestamp.
+node {
+  calculator: "EndLoopNormalizedRectCalculator"
+  input_stream: "ITEM:single_hand_rect_from_landmarks"
+  input_stream: "BATCH_END:single_hand_rect_timestamp"
+  output_stream: "ITERABLE:multi_hand_rects_for_next_frame"
+}
+
+# Filters the input vector of landmarks based on hand presence value for each
+# hand. If the hand presence for hand #i is false, the set of landmarks
+# corresponding to that hand are dropped from the vector.
+node {
+  calculator: "FilterLandmarkListCollectionCalculator"
+  input_stream: "ITERABLE:multi_hand_landmarks"
+  input_stream: "CONDITION:multi_hand_presence"
+  output_stream: "ITERABLE:filtered_multi_hand_landmarks"
+}
+
+# Filters the input vector of NormalizedRect based on hand presence value for
+# each hand. If the hand presence for hand #i is false, the NormalizedRect
+# corresponding to that hand are dropped from the vector.
+node {
+  calculator: "FilterNormalizedRectCollectionCalculator"
+  input_stream: "ITERABLE:multi_hand_rects_for_next_frame"
+  input_stream: "CONDITION:multi_hand_presence"
+  output_stream: "ITERABLE:filtered_multi_hand_rects_for_next_frame"
+}
+```
+
+### Multi-Hand Renderer Subgraph
+
+![multi_hand_renderer_gpu_subgraph.pbtxt](images/mobile/multi_hand_renderer_gpu_subgraph.png)
+
+This graph also uses `BeginLoop` and `EndLoop` calculators to iteratively
+convert a set of hand landmarks per hand instance into corresponding
+`RenderData` objects.
+
+[Source pbtxt file](https://github.com/google/mediapipe/tree/master/mediapipe/graphs/hand_tracking/subgraphs/multi_hand_renderer_gpu.pbtxt)
+
+```bash
+# MediaPipe multi-hand tracking rendering subgraph.
+
+type: "MultiHandRendererSubgraph"
+
+input_stream: "IMAGE:input_image"
+# A vector of NormalizedLandmarks, one for each hand.
+input_stream: "LANDMARKS:multi_hand_landmarks"
+# A vector of NormalizedRect, one for each hand.
+input_stream: "NORM_RECTS:0:multi_palm_rects"
+# A vector of NormalizedRect, one for each hand.
+input_stream: "NORM_RECTS:1:multi_hand_rects"
+# A vector of Detection, one for each hand.
+input_stream: "DETECTIONS:palm_detections"
+output_stream: "IMAGE:output_image"
+
+# Converts detections to drawing primitives for annotation overlay.
+node {
+  calculator: "DetectionsToRenderDataCalculator"
+  input_stream: "DETECTIONS:palm_detections"
+  output_stream: "RENDER_DATA:detection_render_data"
+  node_options: {
+    [type.googleapis.com/mediapipe.DetectionsToRenderDataCalculatorOptions] {
+      thickness: 4.0
+      color { r: 0 g: 255 b: 0 }
+    }
+  }
+}
+
+# Converts normalized rects to drawing primitives for annotation overlay.
+node {
+  calculator: "RectToRenderDataCalculator"
+  input_stream: "NORM_RECTS:multi_hand_rects"
+  output_stream: "RENDER_DATA:multi_hand_rects_render_data"
+  node_options: {
+    [type.googleapis.com/mediapipe.RectToRenderDataCalculatorOptions] {
+      filled: false
+      color { r: 255 g: 0 b: 0 }
+      thickness: 4.0
+    }
+  }
+}
+
+# Converts normalized rects to drawing primitives for annotation overlay.
+node {
+  calculator: "RectToRenderDataCalculator"
+  input_stream: "NORM_RECTS:multi_palm_rects"
+  output_stream: "RENDER_DATA:multi_palm_rects_render_data"
+  node_options: {
+    [type.googleapis.com/mediapipe.RectToRenderDataCalculatorOptions] {
+      filled: false
+      color { r: 125 g: 0 b: 122 }
+      thickness: 4.0
+    }
+  }
+}
+
+# Outputs each element of multi_palm_landmarks at a fake timestamp for the rest
+# of the graph to process. At the end of the loop, outputs the BATCH_END
+# timestamp for downstream calculators to inform them that all elements in the
+# vector have been processed.
+node {
+  calculator: "BeginLoopNormalizedLandmarkListVectorCalculator"
+  input_stream: "ITERABLE:multi_hand_landmarks"
+  output_stream: "ITEM:single_hand_landmarks"
+  output_stream: "BATCH_END:landmark_timestamp"
+}
+
+# Converts landmarks to drawing primitives for annotation overlay.
+node {
+  calculator: "LandmarksToRenderDataCalculator"
+  input_stream: "NORM_LANDMARKS:single_hand_landmarks"
+  output_stream: "RENDER_DATA:single_hand_landmark_render_data"
+  node_options: {
+    [type.googleapis.com/mediapipe.LandmarksToRenderDataCalculatorOptions] {
+      landmark_connections: 0
+      landmark_connections: 1
+      landmark_connections: 1
+      landmark_connections: 2
+      landmark_connections: 2
+      landmark_connections: 3
+      landmark_connections: 3
+      landmark_connections: 4
+      landmark_connections: 0
+      landmark_connections: 5
+      landmark_connections: 5
+      landmark_connections: 6
+      landmark_connections: 6
+      landmark_connections: 7
+      landmark_connections: 7
+      landmark_connections: 8
+      landmark_connections: 5
+      landmark_connections: 9
+      landmark_connections: 9
+      landmark_connections: 10
+      landmark_connections: 10
+      landmark_connections: 11
+      landmark_connections: 11
+      landmark_connections: 12
+      landmark_connections: 9
+      landmark_connections: 13
+      landmark_connections: 13
+      landmark_connections: 14
+      landmark_connections: 14
+      landmark_connections: 15
+      landmark_connections: 15
+      landmark_connections: 16
+      landmark_connections: 13
+      landmark_connections: 17
+      landmark_connections: 0
+      landmark_connections: 17
+      landmark_connections: 17
+      landmark_connections: 18
+      landmark_connections: 18
+      landmark_connections: 19
+      landmark_connections: 19
+      landmark_connections: 20
+      landmark_color { r: 255 g: 0 b: 0 }
+      connection_color { r: 0 g: 255 b: 0 }
+      thickness: 4.0
+    }
+  }
+}
+
+# Collects a RenderData object for each hand into a vector. Upon receiving the
+# BATCH_END timestamp, outputs the vector of RenderData at the BATCH_END
+# timestamp.
+node {
+  calculator: "EndLoopRenderDataCalculator"
+  input_stream: "ITEM:single_hand_landmark_render_data"
+  input_stream: "BATCH_END:landmark_timestamp"
+  output_stream: "ITERABLE:multi_hand_landmarks_render_data"
+}
+
+# Draws annotations and overlays them on top of the input images. Consumes
+# 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: "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"
+}
+```

mediapipe/docs/object_detection_coral_devboard.md

@@ -0,0 +1,20 @@
+## Object Detection on Coral with Webcam
+
+MediaPipe is able to run cross platform across device types like desktop, mobile
+and edge devices. Here is an example of running MediaPipe
+[object detection pipeline](./object_detection_desktop.md) on edge device like
+[Google Coral dev board](https://coral.withgoogle.com/products/dev-board) with
+[Edge TPU](https://cloud.google.com/edge-tpu/). This MediaPipe Coral object
+detection pipeline is running [coral specific quantized version](https://github.com/google/mediapipe/blob/master/mediapipe/examples/coral/models/object-detector-quantized_edgetpu.tflite)
+of the [MediaPipe object detection TFLite model](https://github.com/google/mediapipe/blob/master/mediapipe/models/object_detection_front.tflite)
+accelerated on Edge TPU.
+
+### Cross compilation of MediaPipe Coral binaries in Docker
+
+We recommend building the MediaPipe binaries not on the edge device due to
+limited compute resulting in long build times. Instead, we will build MediaPipe
+binaries using Docker containers on a more powerful host machine. For step by
+step details of cross compiling and running MediaPipe binaries on Coral dev
+board, please refer to [README.md in MediaPipe Coral example folder](https://github.com/google/mediapipe/blob/master/mediapipe/examples/coral/README.md).
+
+![Object Detection running on Coral](images/object_detection_demo_coral.jpg)

mediapipe/docs/object_detection_desktop.md

@@ -54,7 +54,7 @@ below and paste it into
 # MediaPipe graph that performs object detection on desktop with TensorFlow
 # on CPU.
 # Used in the example in
-# mediapipie/examples/desktop/object_detection:object_detection_tensorflow.
+# mediapipe/examples/desktop/object_detection:object_detection_tensorflow.
 
 # Decodes an input video file into images and a video header.
 node {
@@ -218,8 +218,6 @@ $ bazel build -c opt --define MEDIAPIPE_DISABLE_GPU=1 \
 # It should print:
 #Target //mediapipe/examples/desktop/object_detection:object_detection_cpu up-to-date:
 #  bazel-bin/mediapipe/examples/desktop/object_detection/object_detection_cpu
-#INFO: Elapsed time: 16.020s, Forge stats: 13001/13003 actions cached, 2.1s CPU used, 0.0s queue time, 89.0 MB ObjFS output (novel bytes: 88.0 MB), 0.0 MB local output, Critical Path: 10.01s, Remote (41.42% of the time): [queue: 0.00%, setup: 4.21%, process: 12.48%]
-#INFO: Streaming build results to: http://sponge2/1824d4cc-ba63-4350-bdc0-aacbd45b902b
 #INFO: Build completed successfully, 12154 total actions
 
 # This will open up your webcam as long as it is connected and on
@@ -240,7 +238,7 @@ below and paste it into
 # MediaPipe graph that performs object detection on desktop with TensorFlow Lite
 # on CPU.
 # Used in the example in
-# mediapipie/examples/desktop/object_detection:object_detection_tflite.
+# mediapipe/examples/desktop/object_detection:object_detection_tflite.
 
 # max_queue_size limits the number of packets enqueued on any input stream
 # by throttling inputs to the graph. This makes the graph only process one

mediapipe/docs/web.md

@@ -0,0 +1,26 @@
+## MediaPipe on the Web
+
+MediaPipe on the Web is an effort to use [WebAssembly](https://webassembly.org/)
+to bring MediaPipe graphs, calculators, and related technologies to the web. The
+aim is to have all the pieces (ML, rendering, and processing) running directly
+in the browser client-side. The official API is under construction, but the core
+technology has been proven effective, and we can already show interactive
+cross-platform demos using your live webcam.
+
+![image](images/web_effect.gif) ![image](images/web_segmentation.gif)
+
+### Hand Tracking (with and without SIMD support)
+
+For [Chrome Developer Summit 2019](https://developer.chrome.com/devsummit/), we
+used this technology to showcase the potential for performance improvements
+using Chrome experimental [WebAssembly SIMD](https://github.com/WebAssembly/simd)
+support.  Below are two different versions of the
+[MediaPipe Hand Tracking Example](https://mediapipe.readthedocs.io/en/latest/hand_tracking_desktop.html)
+ running on the web:
+
+  1. WebAssembly MVP [demo](https://mediapipe.page.link/cds-ht) running around 5-8 frames per second on Desktop Chrome
+
+  2. WebAssembly SIMD [demo](https://mediapipe.page.link/cds-ht-simd) running around 15-18 frames per second on *Canary* Chrome for Desktop, which must additionally be launched with the option `--js-flags="--experimental-wasm-simd"`
+
+
+NOTE: This page is a work-in-progress. More to come soon!

mediapipe/docs/youtube_8m.md

@@ -64,7 +64,7 @@ videos.
 
 4.  Generate a MediaSequence metadata from the input video.
 
-    Note: the output file is /tmp/mediapipe/metadata.tfrecord
+    Note: the output file is /tmp/mediapipe/metadata.pb
 
     ```bash
     # change clip_end_time_sec to match the length of your video.
@@ -82,8 +82,17 @@ videos.
 
     GLOG_logtostderr=1 bazel-bin/mediapipe/examples/desktop/youtube8m/extract_yt8m_features \
       --calculator_graph_config_file=mediapipe/graphs/youtube8m/feature_extraction.pbtxt \
-      --input_side_packets=input_sequence_example=/tmp/mediapipe/metadata.tfrecord  \
+      --input_side_packets=input_sequence_example=/tmp/mediapipe/metadata.pb  \
-      --output_side_packets=output_sequence_example=/tmp/mediapipe/output.tfrecord
+      --output_side_packets=output_sequence_example=/tmp/mediapipe/features.pb
+    ```
+
+6.  [Optional] Read the features.pb in Python.
+
+    ```
+    import tensorflow as tf
+
+    sequence_example = open('/tmp/mediapipe/features.pb', 'rb').read()
+    print(tf.train.SequenceExample.FromString(sequence_example))
     ```
 
 ## Model Inference for YouTube-8M Challenge
@@ -136,7 +145,7 @@ the inference for both local videos and the dataset
 
 ### Steps to run the YouTube-8M model inference graph with a local video
 
-1.  Make sure you have the output tfrecord from the feature extraction pipeline.
+1.  Make sure you have the features.pb from the feature extraction pipeline.
 
 2.  Copy the baseline model
     [(model card)](https://drive.google.com/file/d/1xTCi9-Nm9dt2KIk8WR0dDFrIssWawyXy/view)
@@ -158,7 +167,7 @@ the inference for both local videos and the dataset
     # overlap is the number of seconds adjacent segments share.
     GLOG_logtostderr=1 bazel-bin/mediapipe/examples/desktop/youtube8m/model_inference \
       --calculator_graph_config_file=mediapipe/graphs/youtube8m/local_video_model_inference.pbtxt \
-      --input_side_packets=input_sequence_example_path=/tmp/mediapipe/output.tfrecord,input_video_path=/absolute/path/to/the/local/video/file,output_video_path=/tmp/mediapipe/annotated_video.mp4,segment_size=5,overlap=4
+      --input_side_packets=input_sequence_example_path=/tmp/mediapipe/features.pb,input_video_path=/absolute/path/to/the/local/video/file,output_video_path=/tmp/mediapipe/annotated_video.mp4,segment_size=5,overlap=4
     ```
 
 4.  View the annotated video.

mediapipe/examples/android/src/java/com/google/mediapipe/apps/edgedetectiongpu/MainActivity.java

@@ -47,7 +47,7 @@ public class MainActivity extends AppCompatActivity {
   static {
     // Load all native libraries needed by the app.
     System.loadLibrary("mediapipe_jni");
-    System.loadLibrary("opencv_java4");
+    System.loadLibrary("opencv_java3");
   }
 
   // {@link SurfaceTexture} where the camera-preview frames can be accessed.

mediapipe/examples/android/src/java/com/google/mediapipe/apps/facedetectioncpu/MainActivity.java

@@ -48,7 +48,7 @@ public class MainActivity extends AppCompatActivity {
   static {
     // Load all native libraries needed by the app.
     System.loadLibrary("mediapipe_jni");
-    System.loadLibrary("opencv_java4");
+    System.loadLibrary("opencv_java3");
   }
 
   // {@link SurfaceTexture} where the camera-preview frames can be accessed.

mediapipe/examples/android/src/java/com/google/mediapipe/apps/facedetectiongpu/MainActivity.java

@@ -48,7 +48,7 @@ public class MainActivity extends AppCompatActivity {
   static {
     // Load all native libraries needed by the app.
     System.loadLibrary("mediapipe_jni");
-    System.loadLibrary("opencv_java4");
+    System.loadLibrary("opencv_java3");
   }
 
   // {@link SurfaceTexture} where the camera-preview frames can be accessed.