Project import generated by Copybara.

GitOrigin-RevId: d0039a576e2db9c0fcefffd26a527df74cbe145b
2020-04-21 22:43:01 -04:00 · 2020-04-16 20:20:07 -07:00 · 2020-04-13 20:15:59 -04:00 · 2020-04-06 21:28:19 -07:00 · 2020-04-06 19:14:13 -07:00 · 2020-03-20 15:28:51 -07:00
308 changed files with 13107 additions and 2589 deletions
@@ -12,9 +12,9 @@ build --copt='-Wno-comment'
 build --copt='-Wno-return-type'
 build --copt='-Wno-unused-local-typedefs'
 build --copt='-Wno-ignored-attributes'
-# Temporarily set the incompatibility flag for Bazel 0.27.0 and above
-build --incompatible_disable_deprecated_attr_params=false
-build --incompatible_depset_is_not_iterable=false
+
+# Tensorflow needs remote repo
+build --experimental_repo_remote_exec

 # Sets the default Apple platform to macOS.
 build --apple_platform_type=macos
@@ -23,6 +23,10 @@ build --apple_platform_type=macos
 build --apple_generate_dsym

 # Android configs.
+# Note: the documentation tells us to use @androidndk//:default_crosstool, but
+# the automatic configuration transition uses //external:android/crosstool.
+# Using it here works and spares us from having two different config_settings
+# for Android.
 build:android --crosstool_top=//external:android/crosstool
 build:android --host_crosstool_top=@bazel_tools//tools/cpp:toolchain
 build:android --linkopt=-landroid
@@ -48,7 +48,7 @@ RUN pip install future
 RUN pip3 install six

 # Install bazel
-ARG BAZEL_VERSION=1.1.0
+ARG BAZEL_VERSION=2.0.0
 RUN mkdir /bazel && \
    wget --no-check-certificate -O /bazel/installer.sh "https://github.com/bazelbuild/bazel/releases/download/${BAZEL_VERSION}/b\
 azel-${BAZEL_VERSION}-installer-linux-x86_64.sh" && \
@@ -9,26 +9,34 @@

 ## ML Solutions in MediaPipe

-* [Face Detection](mediapipe/docs/face_detection_mobile_gpu.md) [[Web Demo]](https://viz.mediapipe.dev/runner/demos/face_detection/face_detection.html)
+* [Face Detection](mediapipe/docs/face_detection_mobile_gpu.md) [(web demo)](https://viz.mediapipe.dev/runner/demos/face_detection/face_detection.html)
+* [Face Mesh](mediapipe/docs/face_mesh_mobile_gpu.md)
+* [Hand Detection](mediapipe/docs/hand_detection_mobile_gpu.md)
+* [Hand Tracking](mediapipe/docs/hand_tracking_mobile_gpu.md) [(web demo)](https://viz.mediapipe.dev/runner/demos/hand_tracking/hand_tracking.html)
 * [Multi-hand Tracking](mediapipe/docs/multi_hand_tracking_mobile_gpu.md)
-* [Hand Tracking](mediapipe/docs/hand_tracking_mobile_gpu.md) [[Web Demo]](https://viz.mediapipe.dev/runner/demos/hand_tracking/hand_tracking.html)
-* [Hair Segmentation](mediapipe/docs/hair_segmentation_mobile_gpu.md) [[Web Demo]](https://viz.mediapipe.dev/runner/demos/hair_segmentation/hair_segmentation.html)
+* [Hair Segmentation](mediapipe/docs/hair_segmentation_mobile_gpu.md) [(web demo)](https://viz.mediapipe.dev/runner/demos/hair_segmentation/hair_segmentation.html)
 * [Object Detection](mediapipe/docs/object_detection_mobile_gpu.md)
 * [Object Detection and Tracking](mediapipe/docs/object_tracking_mobile_gpu.md)
 * [Objectron: 3D Object Detection and Tracking](mediapipe/docs/objectron_mobile_gpu.md)
-* [AutoFlip](mediapipe/docs/autoflip.md)
+* [AutoFlip: Intelligent Video Reframing](mediapipe/docs/autoflip.md)
+* [KNIFT: Template Matching with Neural Image Features](mediapipe/docs/template_matching_mobile_cpu.md)

 ![face_detection](mediapipe/docs/images/mobile/face_detection_android_gpu_small.gif)
-![multi-hand_tracking](mediapipe/docs/images/mobile/multi_hand_tracking_android_gpu_small.gif)
-![hand_tracking](mediapipe/docs/images/mobile/hand_tracking_3d_android_gpu_small.gif)
+![face_mesh](mediapipe/docs/images/mobile/face_mesh_android_gpu_small.gif)
+![hand_tracking](mediapipe/docs/images/mobile/hand_tracking_android_gpu_small.gif)
+![multi-hand_tracking](mediapipe/docs/images/mobile/multi_hand_tracking_3d_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)
 ![object_tracking](mediapipe/docs/images/mobile/object_tracking_android_gpu_small.gif)
+![objectron_shoes](mediapipe/docs/images/mobile/objectron_shoe_android_gpu_small.gif)
+![objectron_chair](mediapipe/docs/images/mobile/objectron_chair_android_gpu_small.gif)
+![template_matching](mediapipe/docs/images/mobile/template_matching_android_cpu_small.gif)

 ## Installation
 Follow these [instructions](mediapipe/docs/install.md).

 ## Getting started
-See mobile, desktop and Google Coral [examples](mediapipe/docs/examples.md).
+See mobile, desktop, web and Google Coral [examples](mediapipe/docs/examples.md).

 Check out some web demos [[Edge detection]](https://viz.mediapipe.dev/runner/demos/edge_detection/edge_detection.html) [[Face detection]](https://viz.mediapipe.dev/runner/demos/face_detection/face_detection.html) [[Hand Tracking]](https://viz.mediapipe.dev/runner/demos/hand_tracking/hand_tracking.html)

@@ -40,10 +48,15 @@ Check out the [Examples page](https://mediapipe.readthedocs.io/en/latest/example
 ## Visualizing MediaPipe graphs
 A web-based visualizer is hosted on [viz.mediapipe.dev](https://viz.mediapipe.dev/). Please also see instructions [here](mediapipe/docs/visualizer.md).

+## Google Open Source Code search
+Search MediaPipe Github repository using [Google Open Source code search](https://t.co/LSZnbMUUnT?amp=1)
+
 ## Videos
 *  [YouTube Channel](https://www.youtube.com/channel/UCObqmpuSMx-usADtL_qdMAw)

 ## Publications
+* [MediaPipe KNIFT: Template-based Feature Matching](https://mediapipe.page.link/knift-blog)
+* [Alfred Camera: Smart camera features using MediaPipe](https://developers.googleblog.com/2020/03/alfred-camera-smart-camera-features-using-mediapipe.html)
 * [MediaPipe Objectron: Real-time 3D Object Detection on Mobile Devices](https://mediapipe.page.link/objectron-aiblog)
 * [AutoFlip: An Open Source Framework for Intelligent Video Reframing](https://mediapipe.page.link/autoflip)
 * [Google Developer Blog: MediaPipe on the Web](https://mediapipe.page.link/webdevblog)
@@ -52,6 +65,7 @@ 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
+* [MediaPipe Seattle Meetup, Google Building Waterside, 13 Feb 2020](https://mediapipe.page.link/seattle2020)
 * [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)
@@ -2,16 +2,15 @@ workspace(name = "mediapipe")

 load("@bazel_tools//tools/build_defs/repo:http.bzl", "http_archive")

-skylib_version = "0.8.0"
+skylib_version = "0.9.0"
 http_archive(
    name = "bazel_skylib",
    type = "tar.gz",
-    url = "https://github.com/bazelbuild/bazel-skylib/releases/download/{}/bazel-skylib.{}.tar.gz".format (skylib_version, skylib_version),
-    sha256 = "2ef429f5d7ce7111263289644d233707dba35e39696377ebab8b0bc701f7818e",
+    url = "https://github.com/bazelbuild/bazel-skylib/releases/download/{}/bazel_skylib-{}.tar.gz".format (skylib_version, skylib_version),
+    sha256 = "1dde365491125a3db70731e25658dfdd3bc5dbdfd11b840b3e987ecf043c7ca0",
 )
 load("@bazel_skylib//lib:versions.bzl", "versions")
-versions.check(minimum_bazel_version = "1.0.0",
-               maximum_bazel_version = "1.2.1")
+versions.check(minimum_bazel_version = "2.0.0")


 # ABSL cpp library lts_2020_02_25
@@ -93,11 +92,13 @@ http_archive(
    build_file = "@//third_party:libyuv.BUILD",
 )

+# Note: protobuf-javalite is no longer released as a separate download, it's included in the main Java download.
+# ...but the Java download is currently broken, so we use the "source" download.
 http_archive(
    name = "com_google_protobuf_javalite",
-    sha256 = "79d102c61e2a479a0b7e5fc167bcfaa4832a0c6aad4a75fa7da0480564931bcc",
-    strip_prefix = "protobuf-384989534b2246d413dbcd750744faab2607b516",
-    urls = ["https://github.com/google/protobuf/archive/384989534b2246d413dbcd750744faab2607b516.zip"],
+    sha256 = "a79d19dcdf9139fa4b81206e318e33d245c4c9da1ffed21c87288ed4380426f9",
+    strip_prefix = "protobuf-3.11.4",
+    urls = ["https://github.com/protocolbuffers/protobuf/archive/v3.11.4.tar.gz"],
 )

 http_archive(
@@ -117,19 +118,18 @@ http_archive(
    ],
 )

-# 2020-02-12
-# The last commit before TensorFlow switched to Bazel 2.0
-_TENSORFLOW_GIT_COMMIT = "77e9ffb9b2bfb1a4f7056e62d84039626923e328"
-_TENSORFLOW_SHA256= "176ccd82f7dd17c5e117b50d353603b129c7a6ccbfebd522ca47cc2a40f33f13"
+# 2020-04-01
+_TENSORFLOW_GIT_COMMIT = "805e47cea96c7e8c6fccf494d40a2392dc99fdd8"
+_TENSORFLOW_SHA256= "9ee3ae604c2e1345ac60345becee6d659364721513f9cb8652eb2e7138320ca5"
 http_archive(
    name = "org_tensorflow",
    urls = [
      "https://mirror.bazel.build/github.com/tensorflow/tensorflow/archive/%s.tar.gz" % _TENSORFLOW_GIT_COMMIT,
      "https://github.com/tensorflow/tensorflow/archive/%s.tar.gz" % _TENSORFLOW_GIT_COMMIT,
    ],
-    # A compatibility patch
    patches = [
-        "@//third_party:org_tensorflow_528e22eae8bf3206189a066032c66e9e5c9b4a61.diff"
+        "@//third_party:org_tensorflow_compatibility_fixes.diff",
+        "@//third_party:org_tensorflow_protobuf_updates.diff",
    ],
    patch_args = [
        "-p1",
@@ -154,10 +154,6 @@ http_archive(
    sha256 = "5ba6d0db4e784621fda44a50c58bb23b0892684692f0c623e2063f9c19f192f1"
 )

-# Please run
-# $ sudo apt-get install libopencv-core-dev libopencv-highgui-dev \
-#                        libopencv-calib3d-dev libopencv-features2d-dev \
-#                        libopencv-imgproc-dev libopencv-video-dev
 new_local_repository(
    name = "linux_opencv",
    build_file = "@//third_party:opencv_linux.BUILD",
@@ -170,7 +166,6 @@ new_local_repository(
    path = "/usr"
 )

-# Please run $ brew install opencv@3
 new_local_repository(
    name = "macos_opencv",
    build_file = "@//third_party:opencv_macos.BUILD",
@@ -203,79 +198,6 @@ http_archive(
    url = "https://github.com/opencv/opencv/releases/download/3.2.0/opencv-3.2.0-ios-framework.zip",
 )

-RULES_JVM_EXTERNAL_TAG = "2.2"
-RULES_JVM_EXTERNAL_SHA = "f1203ce04e232ab6fdd81897cf0ff76f2c04c0741424d192f28e65ae752ce2d6"
-
-http_archive(
-    name = "rules_jvm_external",
-    strip_prefix = "rules_jvm_external-%s" % RULES_JVM_EXTERNAL_TAG,
-    sha256 = RULES_JVM_EXTERNAL_SHA,
-    url = "https://github.com/bazelbuild/rules_jvm_external/archive/%s.zip" % RULES_JVM_EXTERNAL_TAG,
-)
-
-load("@rules_jvm_external//:defs.bzl", "maven_install")
-
-maven_install(
-    artifacts = [
-        "androidx.annotation:annotation:aar:1.1.0",
-        "androidx.appcompat:appcompat:aar:1.1.0-rc01",
-        "androidx.camera:camera-core:aar:1.0.0-alpha06",
-        "androidx.camera:camera-camera2:aar:1.0.0-alpha06",
-        "androidx.constraintlayout:constraintlayout:aar:1.1.3",
-        "androidx.core:core:aar:1.1.0-rc03",
-        "androidx.legacy:legacy-support-v4:aar:1.0.0",
-        "androidx.recyclerview:recyclerview:aar:1.1.0-beta02",
-        "com.google.android.material:material:aar:1.0.0-rc01",
-    ],
-    repositories = [
-        "https://dl.google.com/dl/android/maven2",
-        "https://repo1.maven.org/maven2",
-    ],
-)
-
-maven_server(
-    name = "google_server",
-    url = "https://dl.google.com/dl/android/maven2",
-)
-
-maven_jar(
-    name = "androidx_lifecycle",
-    artifact = "androidx.lifecycle:lifecycle-common:2.0.0",
-    sha1 = "e070ffae07452331bc5684734fce6831d531785c",
-    server = "google_server",
-)
-
-maven_jar(
-    name = "androidx_concurrent_futures",
-    artifact = "androidx.concurrent:concurrent-futures:1.0.0-alpha03",
-    sha1 = "b528df95c7e2fefa2210c0c742bf3e491c1818ae",
-    server = "google_server",
-)
-
-maven_jar(
-    name = "com_google_guava_android",
-    artifact = "com.google.guava:guava:27.0.1-android",
-    sha1 = "b7e1c37f66ef193796ccd7ea6e80c2b05426182d",
-)
-
-maven_jar(
-    name = "com_google_common_flogger",
-    artifact = "com.google.flogger:flogger:0.3.1",
-    sha1 = "585030fe1ec709760cbef997a459729fb965df0e",
-)
-
-maven_jar(
-    name = "com_google_common_flogger_system_backend",
-    artifact = "com.google.flogger:flogger-system-backend:0.3.1",
-    sha1 = "287b569d76abcd82f9de87fe41829fbc7ebd8ac9",
-)
-
-maven_jar(
-    name = "com_google_code_findbugs",
-    artifact = "com.google.code.findbugs:jsr305:3.0.2",
-    sha1 = "25ea2e8b0c338a877313bd4672d3fe056ea78f0d",
-)
-
 # You may run setup_android.sh to install Android SDK and NDK.
 android_ndk_repository(
    name = "androidndk",
@@ -289,9 +211,15 @@ android_sdk_repository(

 http_archive(
    name = "build_bazel_rules_apple",
-    sha256 = "bdc8e66e70b8a75da23b79f1f8c6207356df07d041d96d2189add7ee0780cf4e",
-    strip_prefix = "rules_apple-b869b0d3868d78a1d4ffd866ccb304fb68aa12c3",
-    url = "https://github.com/bazelbuild/rules_apple/archive/b869b0d3868d78a1d4ffd866ccb304fb68aa12c3.tar.gz",
+    sha256 = "7a7afdd4869bb201c9352eed2daf37294d42b093579b70423490c1b4d4f6ce42",
+    url = "https://github.com/bazelbuild/rules_apple/releases/download/0.19.0/rules_apple.0.19.0.tar.gz",
+    patches = [
+        # Bypass checking ios unit test runner when building MP ios applications.
+        "@//third_party:build_bazel_rules_apple_bypass_test_runner_check.diff"
+    ],
+    patch_args = [
+        "-p1",
+    ],
 )

 load(
@@ -325,3 +253,49 @@ http_archive(
    build_file = "@//third_party:google_toolbox_for_mac.BUILD",
 )

+# Maven dependencies.
+
+RULES_JVM_EXTERNAL_TAG = "3.2"
+RULES_JVM_EXTERNAL_SHA = "82262ff4223c5fda6fb7ff8bd63db8131b51b413d26eb49e3131037e79e324af"
+
+http_archive(
+    name = "rules_jvm_external",
+    strip_prefix = "rules_jvm_external-%s" % RULES_JVM_EXTERNAL_TAG,
+    sha256 = RULES_JVM_EXTERNAL_SHA,
+    url = "https://github.com/bazelbuild/rules_jvm_external/archive/%s.zip" % RULES_JVM_EXTERNAL_TAG,
+)
+
+load("@rules_jvm_external//:defs.bzl", "maven_install")
+
+# Important: there can only be one maven_install rule. Add new maven deps here.
+maven_install(
+    name = "maven",
+    artifacts = [
+        "junit:junit:4.12",
+        "androidx.test.espresso:espresso-core:3.1.1",
+        "org.hamcrest:hamcrest-library:1.3",
+        "androidx.concurrent:concurrent-futures:1.0.0-alpha03",
+        "androidx.lifecycle:lifecycle-common:2.2.0",
+        "androidx.annotation:annotation:aar:1.1.0",
+        "androidx.appcompat:appcompat:aar:1.1.0-rc01",
+        "androidx.camera:camera-core:aar:1.0.0-alpha06",
+        "androidx.camera:camera-camera2:aar:1.0.0-alpha06",
+        "androidx.constraintlayout:constraintlayout:aar:1.1.3",
+        "androidx.core:core:aar:1.1.0-rc03",
+        "androidx.legacy:legacy-support-v4:aar:1.0.0",
+        "androidx.recyclerview:recyclerview:aar:1.1.0-beta02",
+        "com.google.android.material:material:aar:1.0.0-rc01",
+        "com.google.code.findbugs:jsr305:3.0.2",
+        "com.google.flogger:flogger-system-backend:0.3.1",
+        "com.google.flogger:flogger:0.3.1",
+        "com.google.guava:guava:27.0.1-android",
+    ],
+    repositories = [
+        "https://jcenter.bintray.com",
+        "https://maven.google.com",
+        "https://dl.google.com/dl/android/maven2",
+        "https://repo1.maven.org/maven2",
+    ],
+    fetch_sources = True,
+    version_conflict_policy = "pinned",
+)
@@ -14,6 +14,9 @@

 licenses(["notice"])  # Apache 2.0

+# Note: yes, these need to use "//external:android/crosstool", not
+# @androidndk//:default_crosstool.
+
 config_setting(
    name = "android",
    values = {"crosstool_top": "//external:android/crosstool"},
@@ -184,23 +184,14 @@ const float SpectrogramCalculator::kLnPowerToDb = 4.342944819032518;
  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"
-           "frame_duration_seconds: "
-        << spectrogram_options.frame_overlap_seconds();
+    // TODO: return an error.
  }
  if (spectrogram_options.frame_overlap_seconds() >=
      spectrogram_options.frame_duration_seconds()) {
-    ::mediapipe::InvalidArgumentErrorBuilder(MEDIAPIPE_LOC)
-        << "Invalid frame_overlap_seconds.\nframe_overlap_seconds: "
-        << spectrogram_options.frame_overlap_seconds()
-        << "\nframe_duration_seconds: "
-        << spectrogram_options.frame_duration_seconds();
+    // TODO: return an error.
  }
  if (spectrogram_options.frame_overlap_seconds() < 0.0) {
-    ::mediapipe::InvalidArgumentErrorBuilder(MEDIAPIPE_LOC)
-        << "Frame_overlap_seconds is < 0.0.\nframe_overlap_seconds: "
-        << spectrogram_options.frame_overlap_seconds();
+    // TODO: return an error.
  }

  TimeSeriesHeader input_header;
@@ -212,9 +203,7 @@ const float SpectrogramCalculator::kLnPowerToDb = 4.342944819032518;

  if (!spectrogram_options.allow_multichannel_input() &&
      num_input_channels_ != 1) {
-    ::mediapipe::InvalidArgumentErrorBuilder(MEDIAPIPE_LOC)
-        << "The current setting only supports single-channel input. Please set "
-           "allow_multichannel_input.\n";
+    // TODO: return an error.
  }

  frame_duration_samples_ =
@@ -293,10 +282,7 @@ const float SpectrogramCalculator::kLnPowerToDb = 4.342944819032518;

  const Matrix& input_stream = cc->Inputs().Index(0).Get<Matrix>();
  if (input_stream.rows() != num_input_channels_) {
-    ::mediapipe::InvalidArgumentErrorBuilder(MEDIAPIPE_LOC)
-        << "Number of input channels do not correspond to the number of rows "
-        << "in the input matrix: " << num_input_channels_ << "channels vs "
-        << input_stream.rows() << " rows";
+    // TODO: return an error.
  }

  cumulative_input_samples_ += input_stream.cols();
@@ -228,7 +228,9 @@ cc_library(
        "//mediapipe/framework:calculator_framework",
        "//mediapipe/framework:collection_item_id",
        "//mediapipe/framework:packet",
+        "//mediapipe/framework/formats:detection_cc_proto",
        "//mediapipe/framework/formats:landmark_cc_proto",
+        "//mediapipe/framework/formats:matrix",
        "//mediapipe/framework/formats:rect_cc_proto",
        "//mediapipe/framework/port:integral_types",
        "//mediapipe/framework/port:ret_check",
@@ -249,12 +251,14 @@ cc_library(
        "//mediapipe/framework:calculator_framework",
        "//mediapipe/framework:collection_item_id",
        "//mediapipe/framework:packet",
+        "//mediapipe/framework/formats:classification_cc_proto",
        "//mediapipe/framework/formats:landmark_cc_proto",
        "//mediapipe/framework/formats:rect_cc_proto",
        "//mediapipe/framework/port:integral_types",
        "//mediapipe/framework/port:ret_check",
        "//mediapipe/framework/port:status",
        "//mediapipe/util:render_data_cc_proto",
+        "@org_tensorflow//tensorflow/lite:framework",
    ],
    alwayslink = 1,
 )
@@ -265,10 +269,11 @@ cc_test(
    deps = [
        ":begin_loop_calculator",
        ":end_loop_calculator",
-        "//mediapipe/calculators/core:packet_cloner_calculator",
+        ":gate_calculator",
        "//mediapipe/framework:calculator_context",
        "//mediapipe/framework:calculator_contract",
        "//mediapipe/framework:calculator_framework",
+        "//mediapipe/framework:packet",
        "//mediapipe/framework/port:gtest_main",
        "//mediapipe/framework/port:integral_types",
        "//mediapipe/framework/port:parse_text_proto",
@@ -334,6 +339,7 @@ cc_library(
    deps = [
        ":clip_vector_size_calculator_cc_proto",
        "//mediapipe/framework:calculator_framework",
+        "//mediapipe/framework/formats:detection_cc_proto",
        "//mediapipe/framework/formats:rect_cc_proto",
        "//mediapipe/framework/port:ret_check",
        "//mediapipe/framework/port:status",
@@ -693,15 +699,17 @@ cc_test(
    name = "previous_loopback_calculator_test",
    srcs = ["previous_loopback_calculator_test.cc"],
    deps = [
+        ":gate_calculator",
+        ":make_pair_calculator",
+        ":pass_through_calculator",
        ":previous_loopback_calculator",
-        "//mediapipe/calculators/core:make_pair_calculator",
-        "//mediapipe/calculators/core:pass_through_calculator",
        "//mediapipe/framework:calculator_framework",
        "//mediapipe/framework:calculator_runner",
        "//mediapipe/framework:timestamp",
        "//mediapipe/framework/port:gtest_main",
        "//mediapipe/framework/port:integral_types",
        "//mediapipe/framework/port:parse_text_proto",
+        "//mediapipe/framework/port:status",
        "//mediapipe/framework/stream_handler:immediate_input_stream_handler",
        "//mediapipe/framework/tool:sink",
        "@com_google_absl//absl/time",
@@ -769,9 +777,11 @@ cc_library(
    visibility = ["//visibility:public"],
    deps = [
        ":split_vector_calculator_cc_proto",
+        "//mediapipe/framework/formats:detection_cc_proto",
        "//mediapipe/framework:calculator_framework",
        "//mediapipe/framework/formats:landmark_cc_proto",
        "//mediapipe/framework/formats:rect_cc_proto",
+        "//mediapipe/framework/formats:matrix",
        "//mediapipe/framework/port:ret_check",
        "//mediapipe/framework/port:status",
        "//mediapipe/util:resource_util",
@@ -805,6 +815,38 @@ cc_test(
    ],
 )

+cc_library(
+    name = "split_normalized_landmark_list_calculator",
+    srcs = ["split_normalized_landmark_list_calculator.cc"],
+    visibility = ["//visibility:public"],
+    deps = [
+        ":split_vector_calculator_cc_proto",
+        "//mediapipe/framework:calculator_framework",
+        "//mediapipe/framework/formats:landmark_cc_proto",
+        "//mediapipe/framework/port:ret_check",
+        "//mediapipe/framework/port:status",
+        "//mediapipe/util:resource_util",
+    ],
+    alwayslink = 1,
+)
+
+cc_test(
+    name = "split_normalized_landmark_list_calculator_test",
+    srcs = ["split_normalized_landmark_list_calculator_test.cc"],
+    deps = [
+        ":split_normalized_landmark_list_calculator",
+        ":split_vector_calculator_cc_proto",
+        "//mediapipe/framework:calculator_framework",
+        "//mediapipe/framework:calculator_runner",
+        "//mediapipe/framework/deps:file_path",
+        "//mediapipe/framework/formats:landmark_cc_proto",
+        "//mediapipe/framework/port:gtest_main",
+        "//mediapipe/framework/port:integral_types",
+        "//mediapipe/framework/port:parse_text_proto",
+        "//mediapipe/framework/tool:validate_type",
+    ],
+)
+
 cc_library(
    name = "dequantize_byte_array_calculator",
    srcs = ["dequantize_byte_array_calculator.cc"],
@@ -20,6 +20,8 @@
 #include "mediapipe/calculators/core/end_loop_calculator.h"
 #include "mediapipe/framework/calculator_contract.h"
 #include "mediapipe/framework/calculator_framework.h"
+#include "mediapipe/framework/packet.h"
+#include "mediapipe/framework/port/gmock.h"
 #include "mediapipe/framework/port/gtest.h"
 #include "mediapipe/framework/port/integral_types.h"
 #include "mediapipe/framework/port/parse_text_proto.h"
@@ -28,6 +30,13 @@
 namespace mediapipe {
 namespace {

+MATCHER_P2(PacketOfIntsEq, timestamp, value, "") {
+  Timestamp actual_timestamp = arg.Timestamp();
+  const auto& actual_value = arg.template Get<std::vector<int>>();
+  return testing::Value(actual_timestamp, testing::Eq(timestamp)) &&
+         testing::Value(actual_value, testing::ElementsAreArray(value));
+}
+
 typedef BeginLoopCalculator<std::vector<int>> BeginLoopIntegerCalculator;
 REGISTER_CALCULATOR(BeginLoopIntegerCalculator);

@@ -59,8 +68,8 @@ REGISTER_CALCULATOR(EndLoopIntegersCalculator);

 class BeginEndLoopCalculatorGraphTest : public ::testing::Test {
 protected:
-  BeginEndLoopCalculatorGraphTest() {
-    graph_config_ = ParseTextProtoOrDie<CalculatorGraphConfig>(
+  void SetUp() override {
+    auto graph_config = ParseTextProtoOrDie<CalculatorGraphConfig>(
        R"(
          num_threads: 4
          input_stream: "ints"
@@ -82,94 +91,222 @@ class BeginEndLoopCalculatorGraphTest : public ::testing::Test {
            output_stream: "ITERABLE:ints_plus_one"
          }
        )");
-    tool::AddVectorSink("ints_plus_one", &graph_config_, &output_packets_);
+    tool::AddVectorSink("ints_plus_one", &graph_config, &output_packets_);
+    MP_ASSERT_OK(graph_.Initialize(graph_config));
+    MP_ASSERT_OK(graph_.StartRun({}));
  }

-  CalculatorGraphConfig graph_config_;
+  void SendPacketOfInts(Timestamp timestamp, std::vector<int> ints) {
+    MP_ASSERT_OK(graph_.AddPacketToInputStream(
+        "ints", MakePacket<std::vector<int>>(std::move(ints)).At(timestamp)));
+  }
+
+  CalculatorGraph graph_;
  std::vector<Packet> output_packets_;
 };

+TEST_F(BeginEndLoopCalculatorGraphTest, InputStreamForIterableIsEmpty) {
+  MP_ASSERT_OK(graph_.WaitUntilIdle());
+
+  // EndLoopCalc will forward the timestamp bound because there are no packets
+  // to process.
+  ASSERT_EQ(0, output_packets_.size());
+
+  MP_ASSERT_OK(graph_.CloseAllPacketSources());
+  MP_ASSERT_OK(graph_.WaitUntilDone());
+}
+
 TEST_F(BeginEndLoopCalculatorGraphTest, SingleEmptyVector) {
-  CalculatorGraph graph;
-  MP_EXPECT_OK(graph.Initialize(graph_config_));
-  MP_EXPECT_OK(graph.StartRun({}));
-  auto input_vector = absl::make_unique<std::vector<int>>();
-  Timestamp input_timestamp = Timestamp(0);
-  MP_ASSERT_OK(graph.AddPacketToInputStream(
-      "ints", Adopt(input_vector.release()).At(input_timestamp)));
-  MP_ASSERT_OK(graph.WaitUntilIdle());
+  SendPacketOfInts(Timestamp(0), {});
+  MP_ASSERT_OK(graph_.WaitUntilIdle());

  // EndLoopCalc will forward the timestamp bound because there are no elements
  // in collection to output.
-  ASSERT_EQ(0, output_packets_.size());
+  EXPECT_TRUE(output_packets_.empty());

-  MP_ASSERT_OK(graph.CloseAllPacketSources());
-  MP_ASSERT_OK(graph.WaitUntilDone());
+  MP_ASSERT_OK(graph_.CloseAllPacketSources());
+  MP_ASSERT_OK(graph_.WaitUntilDone());
 }

 TEST_F(BeginEndLoopCalculatorGraphTest, SingleNonEmptyVector) {
-  CalculatorGraph graph;
-  MP_EXPECT_OK(graph.Initialize(graph_config_));
-  MP_EXPECT_OK(graph.StartRun({}));
-  auto input_vector = absl::make_unique<std::vector<int>>();
-  input_vector->emplace_back(0);
-  input_vector->emplace_back(1);
-  input_vector->emplace_back(2);
  Timestamp input_timestamp = Timestamp(0);
-  MP_ASSERT_OK(graph.AddPacketToInputStream(
-      "ints", Adopt(input_vector.release()).At(input_timestamp)));
-  MP_ASSERT_OK(graph.WaitUntilIdle());
+  SendPacketOfInts(input_timestamp, {0, 1, 2});
+  MP_ASSERT_OK(graph_.WaitUntilIdle());

-  ASSERT_EQ(1, output_packets_.size());
-  EXPECT_EQ(input_timestamp, output_packets_[0].Timestamp());
-  std::vector<int> expected_output_vector = {1, 2, 3};
-  EXPECT_EQ(expected_output_vector, output_packets_[0].Get<std::vector<int>>());
+  EXPECT_THAT(output_packets_,
+              testing::ElementsAre(
+                  PacketOfIntsEq(input_timestamp, std::vector<int>{1, 2, 3})));

-  MP_ASSERT_OK(graph.CloseAllPacketSources());
-  MP_ASSERT_OK(graph.WaitUntilDone());
+  MP_ASSERT_OK(graph_.CloseAllPacketSources());
+  MP_ASSERT_OK(graph_.WaitUntilDone());
 }

 TEST_F(BeginEndLoopCalculatorGraphTest, MultipleVectors) {
-  CalculatorGraph graph;
-  MP_EXPECT_OK(graph.Initialize(graph_config_));
-  MP_EXPECT_OK(graph.StartRun({}));
-
-  auto input_vector0 = absl::make_unique<std::vector<int>>();
-  input_vector0->emplace_back(0);
-  input_vector0->emplace_back(1);
  Timestamp input_timestamp0 = Timestamp(0);
-  MP_ASSERT_OK(graph.AddPacketToInputStream(
-      "ints", Adopt(input_vector0.release()).At(input_timestamp0)));
+  SendPacketOfInts(input_timestamp0, {0, 1});

-  auto input_vector1 = absl::make_unique<std::vector<int>>();
  Timestamp input_timestamp1 = Timestamp(1);
-  MP_ASSERT_OK(graph.AddPacketToInputStream(
-      "ints", Adopt(input_vector1.release()).At(input_timestamp1)));
+  SendPacketOfInts(input_timestamp1, {});

-  auto input_vector2 = absl::make_unique<std::vector<int>>();
-  input_vector2->emplace_back(2);
-  input_vector2->emplace_back(3);
  Timestamp input_timestamp2 = Timestamp(2);
-  MP_ASSERT_OK(graph.AddPacketToInputStream(
-      "ints", Adopt(input_vector2.release()).At(input_timestamp2)));
+  SendPacketOfInts(input_timestamp2, {2, 3});

-  MP_ASSERT_OK(graph.CloseAllPacketSources());
-  MP_ASSERT_OK(graph.WaitUntilDone());
-
-  ASSERT_EQ(2, output_packets_.size());
-
-  EXPECT_EQ(input_timestamp0, output_packets_[0].Timestamp());
-  std::vector<int> expected_output_vector0 = {1, 2};
-  EXPECT_EQ(expected_output_vector0,
-            output_packets_[0].Get<std::vector<int>>());
+  MP_ASSERT_OK(graph_.CloseAllPacketSources());
+  MP_ASSERT_OK(graph_.WaitUntilDone());

  // At input_timestamp1, EndLoopCalc will forward timestamp bound as there are
  // no elements in vector to process.
+  EXPECT_THAT(output_packets_,
+              testing::ElementsAre(
+                  PacketOfIntsEq(input_timestamp0, std::vector<int>{1, 2}),
+                  PacketOfIntsEq(input_timestamp2, std::vector<int>{3, 4})));
+}

-  EXPECT_EQ(input_timestamp2, output_packets_[1].Timestamp());
-  std::vector<int> expected_output_vector2 = {3, 4};
-  EXPECT_EQ(expected_output_vector2,
-            output_packets_[1].Get<std::vector<int>>());
+// Passes non empty vector through or outputs empty vector in case of timestamp
+// bound update.
+class PassThroughOrEmptyVectorCalculator : public CalculatorBase {
+ public:
+  static ::mediapipe::Status GetContract(CalculatorContract* cc) {
+    cc->SetProcessTimestampBounds(true);
+    cc->Inputs().Index(0).Set<std::vector<int>>();
+    cc->Outputs().Index(0).Set<std::vector<int>>();
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Open(CalculatorContext* cc) override {
+    cc->SetOffset(TimestampDiff(0));
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Process(CalculatorContext* cc) override {
+    if (!cc->Inputs().Index(0).IsEmpty()) {
+      cc->Outputs().Index(0).AddPacket(cc->Inputs().Index(0).Value());
+    } else {
+      cc->Outputs().Index(0).AddPacket(
+          MakePacket<std::vector<int>>(std::vector<int>())
+              .At(cc->InputTimestamp()));
+    }
+    return ::mediapipe::OkStatus();
+  }
+};
+
+REGISTER_CALCULATOR(PassThroughOrEmptyVectorCalculator);
+
+class BeginEndLoopCalculatorGraphProcessingEmptyPacketsTest
+    : public ::testing::Test {
+ protected:
+  void SetUp() override {
+    auto graph_config = ParseTextProtoOrDie<CalculatorGraphConfig>(
+        R"(
+          num_threads: 4
+          input_stream: "ints"
+          input_stream: "force_ints_to_be_timestamp_bound_update"
+          node {
+            calculator: "GateCalculator"
+            input_stream: "ints"
+            input_stream: "DISALLOW:force_ints_to_be_timestamp_bound_update"
+            output_stream: "ints_passed_through"
+          }
+          node {
+            calculator: "BeginLoopIntegerCalculator"
+            input_stream: "ITERABLE:ints_passed_through"
+            output_stream: "ITEM:int"
+            output_stream: "BATCH_END:timestamp"
+          }
+          node {
+            calculator: "IncrementCalculator"
+            input_stream: "int"
+            output_stream: "int_plus_one"
+          }
+          node {
+            calculator: "EndLoopIntegersCalculator"
+            input_stream: "ITEM:int_plus_one"
+            input_stream: "BATCH_END:timestamp"
+            output_stream: "ITERABLE:ints_plus_one"
+          }
+          node {
+            calculator: "PassThroughOrEmptyVectorCalculator"
+            input_stream: "ints_plus_one"
+            output_stream: "ints_plus_one_passed_through"
+          }
+        )");
+    tool::AddVectorSink("ints_plus_one_passed_through", &graph_config,
+                        &output_packets_);
+    MP_ASSERT_OK(graph_.Initialize(graph_config));
+    MP_ASSERT_OK(graph_.StartRun({}));
+  }
+
+  void SendPacketOfIntsOrBound(Timestamp timestamp, std::vector<int> ints) {
+    // All "ints" packets which are empty are forced to be just timestamp
+    // bound updates for begin loop calculator.
+    bool force_ints_to_be_timestamp_bound_update = ints.empty();
+    MP_ASSERT_OK(graph_.AddPacketToInputStream(
+        "force_ints_to_be_timestamp_bound_update",
+        MakePacket<bool>(force_ints_to_be_timestamp_bound_update)
+            .At(timestamp)));
+    MP_ASSERT_OK(graph_.AddPacketToInputStream(
+        "ints", MakePacket<std::vector<int>>(std::move(ints)).At(timestamp)));
+  }
+
+  CalculatorGraph graph_;
+  std::vector<Packet> output_packets_;
+};
+
+TEST_F(BeginEndLoopCalculatorGraphProcessingEmptyPacketsTest,
+       SingleEmptyVector) {
+  SendPacketOfIntsOrBound(Timestamp(0), {});
+  MP_ASSERT_OK(graph_.WaitUntilIdle());
+
+  EXPECT_THAT(output_packets_, testing::ElementsAre(PacketOfIntsEq(
+                                   Timestamp(0), std::vector<int>{})));
+
+  MP_ASSERT_OK(graph_.CloseAllPacketSources());
+  MP_ASSERT_OK(graph_.WaitUntilDone());
+}
+
+TEST_F(BeginEndLoopCalculatorGraphProcessingEmptyPacketsTest,
+       SingleNonEmptyVector) {
+  SendPacketOfIntsOrBound(Timestamp(0), {0, 1, 2});
+  MP_ASSERT_OK(graph_.WaitUntilIdle());
+
+  EXPECT_THAT(output_packets_, testing::ElementsAre(PacketOfIntsEq(
+                                   Timestamp(0), std::vector<int>{1, 2, 3})));
+
+  MP_ASSERT_OK(graph_.CloseAllPacketSources());
+  MP_ASSERT_OK(graph_.WaitUntilDone());
+}
+
+TEST_F(BeginEndLoopCalculatorGraphProcessingEmptyPacketsTest, MultipleVectors) {
+  SendPacketOfIntsOrBound(Timestamp(0), {});
+  // Waiting until idle to guarantee all timestamp bound updates are processed
+  // individually. (Timestamp bounds updates occur in the provide config only
+  // if input is an empty vector.)
+  MP_ASSERT_OK(graph_.WaitUntilIdle());
+
+  SendPacketOfIntsOrBound(Timestamp(1), {0, 1});
+  SendPacketOfIntsOrBound(Timestamp(2), {});
+  // Waiting until idle to guarantee all timestamp bound updates are processed
+  // individually. (Timestamp bounds updates occur in the provide config only
+  // if input is an empty vector.)
+  MP_ASSERT_OK(graph_.WaitUntilIdle());
+
+  SendPacketOfIntsOrBound(Timestamp(3), {2, 3});
+  SendPacketOfIntsOrBound(Timestamp(4), {});
+  // Waiting until idle to guarantee all timestamp bound updates are processed
+  // individually. (Timestamp bounds updates occur in the provide config only
+  // if input is an empty vector.)
+  MP_ASSERT_OK(graph_.WaitUntilIdle());
+
+  MP_ASSERT_OK(graph_.CloseAllPacketSources());
+  MP_ASSERT_OK(graph_.WaitUntilDone());
+
+  EXPECT_THAT(
+      output_packets_,
+      testing::ElementsAre(PacketOfIntsEq(Timestamp(0), std::vector<int>{}),
+                           PacketOfIntsEq(Timestamp(1), std::vector<int>{1, 2}),
+                           PacketOfIntsEq(Timestamp(2), std::vector<int>{}),
+                           PacketOfIntsEq(Timestamp(3), std::vector<int>{3, 4}),
+                           PacketOfIntsEq(Timestamp(4), std::vector<int>{})));
 }

 class MultiplierCalculator : public CalculatorBase {
@@ -199,8 +336,8 @@ REGISTER_CALCULATOR(MultiplierCalculator);

 class BeginEndLoopCalculatorGraphWithClonedInputsTest : public ::testing::Test {
 protected:
-  BeginEndLoopCalculatorGraphWithClonedInputsTest() {
-    graph_config_ = ParseTextProtoOrDie<CalculatorGraphConfig>(
+  void SetUp() override {
+    auto graph_config = ParseTextProtoOrDie<CalculatorGraphConfig>(
        R"(
          num_threads: 4
          input_stream: "ints"
@@ -226,109 +363,85 @@ class BeginEndLoopCalculatorGraphWithClonedInputsTest : public ::testing::Test {
            output_stream: "ITERABLE:multiplied_ints"
          }
        )");
-    tool::AddVectorSink("multiplied_ints", &graph_config_, &output_packets_);
+    tool::AddVectorSink("multiplied_ints", &graph_config, &output_packets_);
+    MP_ASSERT_OK(graph_.Initialize(graph_config));
+    MP_ASSERT_OK(graph_.StartRun({}));
  }

-  CalculatorGraphConfig graph_config_;
+  void SendPackets(Timestamp timestamp, int multiplier, std::vector<int> ints) {
+    MP_ASSERT_OK(graph_.AddPacketToInputStream(
+        "ints", MakePacket<std::vector<int>>(std::move(ints)).At(timestamp)));
+    MP_ASSERT_OK(graph_.AddPacketToInputStream(
+        "multiplier", MakePacket<int>(multiplier).At(timestamp)));
+  }
+
+  void SendMultiplier(Timestamp timestamp, int multiplier) {
+    MP_ASSERT_OK(graph_.AddPacketToInputStream(
+        "multiplier", MakePacket<int>(multiplier).At(timestamp)));
+  }
+
+  CalculatorGraph graph_;
  std::vector<Packet> output_packets_;
 };

-TEST_F(BeginEndLoopCalculatorGraphWithClonedInputsTest, SingleEmptyVector) {
-  CalculatorGraph graph;
-  MP_EXPECT_OK(graph.Initialize(graph_config_));
-  MP_EXPECT_OK(graph.StartRun({}));
-  auto input_vector = absl::make_unique<std::vector<int>>();
+TEST_F(BeginEndLoopCalculatorGraphWithClonedInputsTest,
+       InputStreamForIterableIsEmpty) {
  Timestamp input_timestamp = Timestamp(42);
-  MP_ASSERT_OK(graph.AddPacketToInputStream(
-      "ints", Adopt(input_vector.release()).At(input_timestamp)));
-  auto multiplier = absl::make_unique<int>(2);
-  MP_ASSERT_OK(graph.AddPacketToInputStream(
-      "multiplier", Adopt(multiplier.release()).At(input_timestamp)));
-  MP_ASSERT_OK(graph.WaitUntilIdle());
+  SendMultiplier(input_timestamp, /*multiplier=*/2);
+  MP_ASSERT_OK(graph_.WaitUntilIdle());
+
+  // EndLoopCalc will forward the timestamp bound because there are no packets
+  // to process.
+  ASSERT_EQ(0, output_packets_.size());
+
+  MP_ASSERT_OK(graph_.CloseAllPacketSources());
+  MP_ASSERT_OK(graph_.WaitUntilDone());
+}
+
+TEST_F(BeginEndLoopCalculatorGraphWithClonedInputsTest, SingleEmptyVector) {
+  SendPackets(Timestamp(0), /*multiplier=*/2, /*ints=*/{});
+  MP_ASSERT_OK(graph_.WaitUntilIdle());

  // EndLoopCalc will forward the timestamp bound because there are no elements
  // in collection to output.
-  ASSERT_EQ(0, output_packets_.size());
+  EXPECT_TRUE(output_packets_.empty());

-  MP_ASSERT_OK(graph.CloseAllPacketSources());
-  MP_ASSERT_OK(graph.WaitUntilDone());
+  MP_ASSERT_OK(graph_.CloseAllPacketSources());
+  MP_ASSERT_OK(graph_.WaitUntilDone());
 }

 TEST_F(BeginEndLoopCalculatorGraphWithClonedInputsTest, SingleNonEmptyVector) {
-  CalculatorGraph graph;
-  MP_EXPECT_OK(graph.Initialize(graph_config_));
-  MP_EXPECT_OK(graph.StartRun({}));
-  auto input_vector = absl::make_unique<std::vector<int>>();
-  input_vector->emplace_back(0);
-  input_vector->emplace_back(1);
-  input_vector->emplace_back(2);
  Timestamp input_timestamp = Timestamp(42);
-  MP_ASSERT_OK(graph.AddPacketToInputStream(
-      "ints", Adopt(input_vector.release()).At(input_timestamp)));
-  auto multiplier = absl::make_unique<int>(2);
-  MP_ASSERT_OK(graph.AddPacketToInputStream(
-      "multiplier", Adopt(multiplier.release()).At(input_timestamp)));
-  MP_ASSERT_OK(graph.WaitUntilIdle());
+  SendPackets(input_timestamp, /*multiplier=*/2, /*ints=*/{0, 1, 2});
+  MP_ASSERT_OK(graph_.WaitUntilIdle());

-  ASSERT_EQ(1, output_packets_.size());
-  EXPECT_EQ(input_timestamp, output_packets_[0].Timestamp());
-  std::vector<int> expected_output_vector = {0, 2, 4};
-  EXPECT_EQ(expected_output_vector, output_packets_[0].Get<std::vector<int>>());
+  EXPECT_THAT(output_packets_,
+              testing::ElementsAre(
+                  PacketOfIntsEq(input_timestamp, std::vector<int>{0, 2, 4})));

-  MP_ASSERT_OK(graph.CloseAllPacketSources());
-  MP_ASSERT_OK(graph.WaitUntilDone());
+  MP_ASSERT_OK(graph_.CloseAllPacketSources());
+  MP_ASSERT_OK(graph_.WaitUntilDone());
 }

 TEST_F(BeginEndLoopCalculatorGraphWithClonedInputsTest, MultipleVectors) {
-  CalculatorGraph graph;
-  MP_EXPECT_OK(graph.Initialize(graph_config_));
-  MP_EXPECT_OK(graph.StartRun({}));
-
-  auto input_vector0 = absl::make_unique<std::vector<int>>();
-  input_vector0->emplace_back(0);
-  input_vector0->emplace_back(1);
  Timestamp input_timestamp0 = Timestamp(42);
-  MP_ASSERT_OK(graph.AddPacketToInputStream(
-      "ints", Adopt(input_vector0.release()).At(input_timestamp0)));
-  auto multiplier0 = absl::make_unique<int>(2);
-  MP_ASSERT_OK(graph.AddPacketToInputStream(
-      "multiplier", Adopt(multiplier0.release()).At(input_timestamp0)));
+  SendPackets(input_timestamp0, /*multiplier=*/2, /*ints=*/{0, 1});

-  auto input_vector1 = absl::make_unique<std::vector<int>>();
  Timestamp input_timestamp1 = Timestamp(43);
-  MP_ASSERT_OK(graph.AddPacketToInputStream(
-      "ints", Adopt(input_vector1.release()).At(input_timestamp1)));
-  auto multiplier1 = absl::make_unique<int>(2);
-  MP_ASSERT_OK(graph.AddPacketToInputStream(
-      "multiplier", Adopt(multiplier1.release()).At(input_timestamp1)));
+  SendPackets(input_timestamp1, /*multiplier=*/2, /*ints=*/{});

-  auto input_vector2 = absl::make_unique<std::vector<int>>();
-  input_vector2->emplace_back(2);
-  input_vector2->emplace_back(3);
  Timestamp input_timestamp2 = Timestamp(44);
-  MP_ASSERT_OK(graph.AddPacketToInputStream(
-      "ints", Adopt(input_vector2.release()).At(input_timestamp2)));
-  auto multiplier2 = absl::make_unique<int>(3);
-  MP_ASSERT_OK(graph.AddPacketToInputStream(
-      "multiplier", Adopt(multiplier2.release()).At(input_timestamp2)));
+  SendPackets(input_timestamp2, /*multiplier=*/3, /*ints=*/{2, 3});

-  MP_ASSERT_OK(graph.CloseAllPacketSources());
-  MP_ASSERT_OK(graph.WaitUntilDone());
-
-  ASSERT_EQ(2, output_packets_.size());
-
-  EXPECT_EQ(input_timestamp0, output_packets_[0].Timestamp());
-  std::vector<int> expected_output_vector0 = {0, 2};
-  EXPECT_EQ(expected_output_vector0,
-            output_packets_[0].Get<std::vector<int>>());
+  MP_ASSERT_OK(graph_.CloseAllPacketSources());
+  MP_ASSERT_OK(graph_.WaitUntilDone());

  // At input_timestamp1, EndLoopCalc will forward timestamp bound as there are
  // no elements in vector to process.
-
-  EXPECT_EQ(input_timestamp2, output_packets_[1].Timestamp());
-  std::vector<int> expected_output_vector2 = {6, 9};
-  EXPECT_EQ(expected_output_vector2,
-            output_packets_[1].Get<std::vector<int>>());
+  EXPECT_THAT(output_packets_,
+              testing::ElementsAre(
+                  PacketOfIntsEq(input_timestamp0, std::vector<int>{0, 2}),
+                  PacketOfIntsEq(input_timestamp2, std::vector<int>{6, 9})));
 }

 }  // namespace
@@ -16,7 +16,9 @@

 #include <vector>

+#include "mediapipe/framework/formats/detection.pb.h"
 #include "mediapipe/framework/formats/landmark.pb.h"
+#include "mediapipe/framework/formats/matrix.h"
 #include "mediapipe/framework/formats/rect.pb.h"

 namespace mediapipe {
@@ -31,4 +33,13 @@ typedef BeginLoopCalculator<std::vector<::mediapipe::NormalizedRect>>
    BeginLoopNormalizedRectCalculator;
 REGISTER_CALCULATOR(BeginLoopNormalizedRectCalculator);

+// A calculator to process std::vector<Detection>.
+typedef BeginLoopCalculator<std::vector<::mediapipe::Detection>>
+    BeginLoopDetectionCalculator;
+REGISTER_CALCULATOR(BeginLoopDetectionCalculator);
+
+// A calculator to process std::vector<Matrix>.
+typedef BeginLoopCalculator<std::vector<Matrix>> BeginLoopMatrixCalculator;
+REGISTER_CALCULATOR(BeginLoopMatrixCalculator);
+
 }  // namespace mediapipe
@@ -52,20 +52,28 @@ namespace mediapipe {
 //   output_stream: "OUTPUT:aggregated_result"     # IterableU @ext_ts
 // }
 //
-// BeginLoopCalculator accepts an optional input stream tagged with "TICK"
-// which if non-empty, wakes up the calculator and calls
-// BeginLoopCalculator::Process(). Input streams tagged with "CLONE" are cloned
-// to the corresponding output streams at loop timestamps. This ensures that a
-// MediaPipe graph or sub-graph can run multiple times, once per element in the
-// "ITERABLE" for each pakcet clone of the packets in the "CLONE" input streams.
+// Input streams tagged with "CLONE" are cloned to the corresponding output
+// streams at loop timestamps. This ensures that a MediaPipe graph or sub-graph
+// can run multiple times, once per element in the "ITERABLE" for each pakcet
+// clone of the packets in the "CLONE" input streams.
 template <typename IterableT>
 class BeginLoopCalculator : public CalculatorBase {
  using ItemT = typename IterableT::value_type;

 public:
  static ::mediapipe::Status GetContract(CalculatorContract* cc) {
+    // The below enables processing of timestamp bound updates, and that enables
+    // correct timestamp propagation by the companion EndLoopCalculator.
+    //
+    // For instance, Process() function will be still invoked even if upstream
+    // calculator has updated timestamp bound for ITERABLE input instead of
+    // providing actual value.
+    cc->SetProcessTimestampBounds(true);
+
    // A non-empty packet in the optional "TICK" input stream wakes up the
    // calculator.
+    // DEPRECATED as timestamp bound updates are processed by default in this
+    // calculator.
    if (cc->Inputs().HasTag("TICK")) {
      cc->Inputs().Tag("TICK").SetAny();
    }
@@ -17,6 +17,7 @@
 #include <vector>

 #include "mediapipe/framework/calculator_framework.h"
+#include "mediapipe/framework/formats/detection.pb.h"
 #include "mediapipe/framework/formats/rect.pb.h"

 namespace mediapipe {
@@ -25,4 +26,8 @@ typedef ClipVectorSizeCalculator<::mediapipe::NormalizedRect>
    ClipNormalizedRectVectorSizeCalculator;
 REGISTER_CALCULATOR(ClipNormalizedRectVectorSizeCalculator);

+typedef ClipVectorSizeCalculator<::mediapipe::Detection>
+    ClipDetectionVectorSizeCalculator;
+REGISTER_CALCULATOR(ClipDetectionVectorSizeCalculator);
+
 }  // namespace mediapipe
@@ -51,8 +51,8 @@ namespace mediapipe {
 class ConstantSidePacketCalculator : public CalculatorBase {
 public:
  static ::mediapipe::Status GetContract(CalculatorContract* cc) {
-    const auto& options = cc->Options().GetExtension(
-        ::mediapipe::ConstantSidePacketCalculatorOptions::ext);
+    const auto& options =
+        cc->Options<::mediapipe::ConstantSidePacketCalculatorOptions>();
    RET_CHECK_EQ(cc->OutputSidePackets().NumEntries(kPacketTag),
                 options.packet_size())
        << "Number of output side packets has to be same as number of packets "
@@ -80,8 +80,8 @@ class ConstantSidePacketCalculator : public CalculatorBase {
  }

  ::mediapipe::Status Open(CalculatorContext* cc) override {
-    const auto& options = cc->Options().GetExtension(
-        ::mediapipe::ConstantSidePacketCalculatorOptions::ext);
+    const auto& options =
+        cc->Options<::mediapipe::ConstantSidePacketCalculatorOptions>();
    int index = 0;
    for (CollectionItemId id = cc->OutputSidePackets().BeginId(kPacketTag);
         id != cc->OutputSidePackets().EndId(kPacketTag); ++id, ++index) {
@@ -16,9 +16,11 @@

 #include <vector>

+#include "mediapipe/framework/formats/classification.pb.h"
 #include "mediapipe/framework/formats/landmark.pb.h"
 #include "mediapipe/framework/formats/rect.pb.h"
 #include "mediapipe/util/render_data.pb.h"
+#include "tensorflow/lite/interpreter.h"

 namespace mediapipe {

@@ -37,4 +39,11 @@ typedef EndLoopCalculator<std::vector<::mediapipe::RenderData>>
    EndLoopRenderDataCalculator;
 REGISTER_CALCULATOR(EndLoopRenderDataCalculator);

+typedef EndLoopCalculator<std::vector<::mediapipe::ClassificationList>>
+    EndLoopClassificationListCalculator;
+REGISTER_CALCULATOR(EndLoopClassificationListCalculator);
+
+typedef EndLoopCalculator<std::vector<TfLiteTensor>> EndLoopTensorCalculator;
+REGISTER_CALCULATOR(EndLoopTensorCalculator);
+
 }  // namespace mediapipe
@@ -25,13 +25,17 @@ namespace mediapipe {
 // together with some previous output.
 //
 // For the first packet that arrives on the MAIN input, the timestamp bound is
-// advanced on the output. Downstream calculators will see this as an empty
+// advanced on the PREV_LOOP. Downstream calculators will see this as an empty
 // packet. This way they are not kept waiting for the previous output, which
 // for the first iteration does not exist.
 //
-// Thereafter, each packet received on MAIN is matched with a packet received
-// on LOOP; the LOOP packet's timestamp is changed to that of the MAIN packet,
-// and it is output on PREV_LOOP.
+// Thereafter,
+// - Each non-empty MAIN packet results in:
+//   a) a PREV_LOOP packet with contents of the LOOP packet received at the
+//      timestamp of the previous non-empty MAIN packet
+//   b) or in a PREV_LOOP timestamp bound update if the LOOP packet was empty.
+// - Each empty MAIN packet indicating timestamp bound update results in a
+//   PREV_LOOP timestamp bound update.
 //
 // Example config:
 // node {
@@ -56,83 +60,115 @@ class PreviousLoopbackCalculator : public CalculatorBase {
    // TODO: an optional PREV_TIMESTAMP output could be added to
    // carry the original timestamp of the packet on PREV_LOOP.
    cc->SetInputStreamHandler("ImmediateInputStreamHandler");
+    // Process() function is invoked in response to MAIN/LOOP stream timestamp
+    // bound updates.
+    cc->SetProcessTimestampBounds(true);
    return ::mediapipe::OkStatus();
  }

  ::mediapipe::Status Open(CalculatorContext* cc) final {
    main_id_ = cc->Inputs().GetId("MAIN", 0);
    loop_id_ = cc->Inputs().GetId("LOOP", 0);
-    loop_out_id_ = cc->Outputs().GetId("PREV_LOOP", 0);
+    prev_loop_id_ = cc->Outputs().GetId("PREV_LOOP", 0);
    cc->Outputs()
-        .Get(loop_out_id_)
+        .Get(prev_loop_id_)
        .SetHeader(cc->Inputs().Get(loop_id_).Header());
-
-    // Use an empty packet for the first round, since there is no previous
-    // output.
-    loopback_packets_.push_back({});
-
    return ::mediapipe::OkStatus();
  }

  ::mediapipe::Status Process(CalculatorContext* cc) final {
-    Packet& main_packet = cc->Inputs().Get(main_id_).Value();
-    if (!main_packet.IsEmpty()) {
-      main_ts_.push_back(main_packet.Timestamp());
-    }
-    Packet& loopback_packet = cc->Inputs().Get(loop_id_).Value();
-    if (!loopback_packet.IsEmpty()) {
-      loopback_packets_.push_back(loopback_packet);
-      while (!main_ts_.empty() &&
-             main_ts_.front() <= loopback_packets_.front().Timestamp()) {
-        main_ts_.pop_front();
-      }
-    }
-    auto& loop_out = cc->Outputs().Get(loop_out_id_);
+    // Non-empty packets and empty packets indicating timestamp bound updates
+    // are guaranteed to have timestamps greater than timestamps of previous
+    // packets within the same stream. Calculator tracks and operates on such
+    // packets.

-    while (!main_ts_.empty() && !loopback_packets_.empty()) {
-      Timestamp main_timestamp = main_ts_.front();
-      main_ts_.pop_front();
-      Packet previous_loopback = loopback_packets_.front().At(main_timestamp);
-      loopback_packets_.pop_front();
-
-      if (previous_loopback.IsEmpty()) {
-        // TODO: SetCompleteTimestampBound would be more useful.
-        loop_out.SetNextTimestampBound(main_timestamp + 1);
+    const Packet& main_packet = cc->Inputs().Get(main_id_).Value();
+    if (prev_main_ts_ < main_packet.Timestamp()) {
+      Timestamp loop_timestamp;
+      if (!main_packet.IsEmpty()) {
+        loop_timestamp = prev_non_empty_main_ts_;
+        prev_non_empty_main_ts_ = main_packet.Timestamp();
      } else {
-        loop_out.AddPacket(std::move(previous_loopback));
+        // Calculator advances PREV_LOOP timestamp bound in response to empty
+        // MAIN packet, hence not caring about corresponding loop packet.
+        loop_timestamp = Timestamp::Unset();
+      }
+      main_packet_specs_.push_back({.timestamp = main_packet.Timestamp(),
+                                    .loop_timestamp = loop_timestamp});
+      prev_main_ts_ = main_packet.Timestamp();
+    }
+
+    const Packet& loop_packet = cc->Inputs().Get(loop_id_).Value();
+    if (prev_loop_ts_ < loop_packet.Timestamp()) {
+      loop_packets_.push_back(loop_packet);
+      prev_loop_ts_ = loop_packet.Timestamp();
+    }
+
+    auto& prev_loop = cc->Outputs().Get(prev_loop_id_);
+    while (!main_packet_specs_.empty() && !loop_packets_.empty()) {
+      // The earliest MAIN packet.
+      const MainPacketSpec& main_spec = main_packet_specs_.front();
+      // The earliest LOOP packet.
+      const Packet& loop_candidate = loop_packets_.front();
+      // Match LOOP and MAIN packets.
+      if (main_spec.loop_timestamp < loop_candidate.Timestamp()) {
+        // No LOOP packet can match the MAIN packet under review.
+        prev_loop.SetNextTimestampBound(main_spec.timestamp + 1);
+        main_packet_specs_.pop_front();
+      } else if (main_spec.loop_timestamp > loop_candidate.Timestamp()) {
+        // No MAIN packet can match the LOOP packet under review.
+        loop_packets_.pop_front();
+      } else {
+        // Exact match found.
+        if (loop_candidate.IsEmpty()) {
+          // However, LOOP packet is empty.
+          prev_loop.SetNextTimestampBound(main_spec.timestamp + 1);
+        } else {
+          prev_loop.AddPacket(loop_candidate.At(main_spec.timestamp));
+        }
+        loop_packets_.pop_front();
+        main_packet_specs_.pop_front();
      }
    }

-    // In case of an empty loopback input, the next timestamp bound for
-    // loopback input is the loopback timestamp + 1.  The next timestamp bound
-    // for output is set and the main_ts_ vector is truncated accordingly.
-    if (loopback_packet.IsEmpty() &&
-        loopback_packet.Timestamp() != Timestamp::Unstarted()) {
-      Timestamp loopback_bound =
-          loopback_packet.Timestamp().NextAllowedInStream();
-      while (!main_ts_.empty() && main_ts_.front() <= loopback_bound) {
-        main_ts_.pop_front();
-      }
-      if (main_ts_.empty()) {
-        loop_out.SetNextTimestampBound(loopback_bound.NextAllowedInStream());
-      }
-    }
-    if (!main_ts_.empty()) {
-      loop_out.SetNextTimestampBound(main_ts_.front());
-    }
-    if (cc->Inputs().Get(main_id_).IsDone() && main_ts_.empty()) {
-      loop_out.Close();
+    if (main_packet_specs_.empty() && cc->Inputs().Get(main_id_).IsDone()) {
+      prev_loop.Close();
    }
    return ::mediapipe::OkStatus();
  }

 private:
+  struct MainPacketSpec {
+    Timestamp timestamp;
+    // Expected timestamp of the packet from LOOP stream that corresponds to the
+    // packet from MAIN stream descirbed by this spec.
+    Timestamp loop_timestamp;
+  };
+
  CollectionItemId main_id_;
  CollectionItemId loop_id_;
-  CollectionItemId loop_out_id_;
+  CollectionItemId prev_loop_id_;

-  std::deque<Timestamp> main_ts_;
-  std::deque<Packet> loopback_packets_;
+  // Contains specs for MAIN packets which only can be:
+  // - non-empty packets
+  // - empty packets indicating timestamp bound updates
+  //
+  // Sorted according to packet timestamps.
+  std::deque<MainPacketSpec> main_packet_specs_;
+  Timestamp prev_main_ts_ = Timestamp::Unstarted();
+  Timestamp prev_non_empty_main_ts_ = Timestamp::Unstarted();
+
+  // Contains LOOP packets which only can be:
+  // - the very first empty packet
+  // - non empty packets
+  // - empty packets indicating timestamp bound updates
+  //
+  // Sorted according to packet timestamps.
+  std::deque<Packet> loop_packets_;
+  // Using "Timestamp::Unset" instead of "Timestamp::Unstarted" in order to
+  // allow addition of the very first empty packet (which doesn't indicate
+  // timestamp bound change necessarily).
+  Timestamp prev_loop_ts_ = Timestamp::Unset();
 };
 REGISTER_CALCULATOR(PreviousLoopbackCalculator);

@@ -13,6 +13,7 @@
 // limitations under the License.

 #include <algorithm>
+#include <functional>
 #include <memory>
 #include <string>
 #include <vector>
@@ -25,12 +26,17 @@
 #include "mediapipe/framework/port/gtest.h"
 #include "mediapipe/framework/port/integral_types.h"
 #include "mediapipe/framework/port/parse_text_proto.h"
+#include "mediapipe/framework/port/status.h"
 #include "mediapipe/framework/port/status_matchers.h"
 #include "mediapipe/framework/timestamp.h"
 #include "mediapipe/framework/tool/sink.h"

 namespace mediapipe {

+using ::testing::ElementsAre;
+using ::testing::Eq;
+using ::testing::Pair;
+using ::testing::Value;
 namespace {

 // Returns the timestamp values for a vector of Packets.
@@ -43,6 +49,23 @@ std::vector<int64> TimestampValues(const std::vector<Packet>& packets) {
  return result;
 }

+MATCHER(EmptyPacket, negation ? "isn't empty" : "is empty") {
+  if (arg.IsEmpty()) {
+    return true;
+  }
+  return false;
+}
+
+MATCHER_P(IntPacket, value, "") {
+  return Value(arg.template Get<int>(), Eq(value));
+}
+
+MATCHER_P2(PairPacket, timestamp, pair, "") {
+  Timestamp actual_timestamp = arg.Timestamp();
+  const auto& actual_pair = arg.template Get<std::pair<Packet, Packet>>();
+  return Value(actual_timestamp, Eq(timestamp)) && Value(actual_pair, pair);
+}
+
 TEST(PreviousLoopbackCalculator, CorrectTimestamps) {
  std::vector<Packet> in_prev;
  CalculatorGraphConfig graph_config_ =
@@ -81,32 +104,30 @@ TEST(PreviousLoopbackCalculator, CorrectTimestamps) {
    MP_EXPECT_OK(graph_.AddPacketToInputStream(
        input_name, MakePacket<int>(n).At(Timestamp(n))));
  };
-  auto pair_values = [](const Packet& packet) {
-    auto pair = packet.Get<std::pair<Packet, Packet>>();
-    int first = pair.first.IsEmpty() ? -1 : pair.first.Get<int>();
-    int second = pair.second.IsEmpty() ? -1 : pair.second.Get<int>();
-    return std::make_pair(first, second);
-  };

  send_packet("in", 1);
  MP_EXPECT_OK(graph_.WaitUntilIdle());
-  EXPECT_EQ(TimestampValues(in_prev), (std::vector<int64>{1}));
-  EXPECT_EQ(pair_values(in_prev.back()), std::make_pair(1, -1));
+  EXPECT_THAT(TimestampValues(in_prev), ElementsAre(1));
+  EXPECT_THAT(in_prev.back(),
+              PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())));

  send_packet("in", 2);
  MP_EXPECT_OK(graph_.WaitUntilIdle());
-  EXPECT_EQ(TimestampValues(in_prev), (std::vector<int64>{1, 2}));
-  EXPECT_EQ(pair_values(in_prev.back()), std::make_pair(2, 1));
+  EXPECT_THAT(TimestampValues(in_prev), ElementsAre(1, 2));
+  EXPECT_THAT(in_prev.back(),
+              PairPacket(Timestamp(2), Pair(IntPacket(2), IntPacket(1))));

  send_packet("in", 5);
  MP_EXPECT_OK(graph_.WaitUntilIdle());
-  EXPECT_EQ(TimestampValues(in_prev), (std::vector<int64>{1, 2, 5}));
-  EXPECT_EQ(pair_values(in_prev.back()), std::make_pair(5, 2));
+  EXPECT_THAT(TimestampValues(in_prev), ElementsAre(1, 2, 5));
+  EXPECT_THAT(in_prev.back(),
+              PairPacket(Timestamp(5), Pair(IntPacket(5), IntPacket(2))));

  send_packet("in", 15);
  MP_EXPECT_OK(graph_.WaitUntilIdle());
-  EXPECT_EQ(TimestampValues(in_prev), (std::vector<int64>{1, 2, 5, 15}));
-  EXPECT_EQ(pair_values(in_prev.back()), std::make_pair(15, 5));
+  EXPECT_THAT(TimestampValues(in_prev), ElementsAre(1, 2, 5, 15));
+  EXPECT_THAT(in_prev.back(),
+              PairPacket(Timestamp(15), Pair(IntPacket(15), IntPacket(5))));

  MP_EXPECT_OK(graph_.CloseAllInputStreams());
  MP_EXPECT_OK(graph_.WaitUntilDone());
@@ -185,24 +206,24 @@ TEST(PreviousLoopbackCalculator, ClosesCorrectly) {

  send_packet("in", 1);
  MP_EXPECT_OK(graph_.WaitUntilIdle());
-  EXPECT_EQ(TimestampValues(outputs), (std::vector<int64>{1}));
+  EXPECT_THAT(TimestampValues(outputs), ElementsAre(1));

  send_packet("in", 2);
  MP_EXPECT_OK(graph_.WaitUntilIdle());
-  EXPECT_EQ(TimestampValues(outputs), (std::vector<int64>{1, 2}));
+  EXPECT_THAT(TimestampValues(outputs), ElementsAre(1, 2));

  send_packet("in", 5);
  MP_EXPECT_OK(graph_.WaitUntilIdle());
-  EXPECT_EQ(TimestampValues(outputs), (std::vector<int64>{1, 2, 5}));
+  EXPECT_THAT(TimestampValues(outputs), ElementsAre(1, 2, 5));

  send_packet("in", 15);
  MP_EXPECT_OK(graph_.WaitUntilIdle());
-  EXPECT_EQ(TimestampValues(outputs), (std::vector<int64>{1, 2, 5, 15}));
+  EXPECT_THAT(TimestampValues(outputs), ElementsAre(1, 2, 5, 15));

  MP_EXPECT_OK(graph_.CloseAllInputStreams());
  MP_EXPECT_OK(graph_.WaitUntilIdle());
-  EXPECT_EQ(TimestampValues(outputs),
-            (std::vector<int64>{1, 2, 5, 15, Timestamp::Max().Value()}));
+  EXPECT_THAT(TimestampValues(outputs),
+              ElementsAre(1, 2, 5, 15, Timestamp::Max().Value()));

  MP_EXPECT_OK(graph_.WaitUntilDone());
 }
@@ -247,16 +268,12 @@ TEST(PreviousLoopbackCalculator, EmptyLoopForever) {
        input_name, MakePacket<int>(n).At(Timestamp(n))));
  };

-  send_packet("in", 0);
-  MP_EXPECT_OK(graph_.WaitUntilIdle());
-  EXPECT_EQ(TimestampValues(outputs), (std::vector<int64>{0}));
-
-  for (int main_ts = 1; main_ts < 50; ++main_ts) {
+  for (int main_ts = 0; main_ts < 50; ++main_ts) {
    send_packet("in", main_ts);
    MP_EXPECT_OK(graph_.WaitUntilIdle());
    std::vector<int64> ts_values = TimestampValues(outputs);
    EXPECT_EQ(ts_values.size(), main_ts + 1);
-    for (int j = 0; j < main_ts; ++j) {
+    for (int j = 0; j < main_ts + 1; ++j) {
      EXPECT_EQ(ts_values[j], j);
    }
  }
@@ -266,5 +283,487 @@ TEST(PreviousLoopbackCalculator, EmptyLoopForever) {
  MP_EXPECT_OK(graph_.WaitUntilDone());
 }

+class PreviousLoopbackCalculatorProcessingTimestampsTest
+    : public testing::Test {
+ protected:
+  void SetUp() override {
+    CalculatorGraphConfig graph_config =
+        ParseTextProtoOrDie<CalculatorGraphConfig>(R"(
+          input_stream: 'input'
+          input_stream: 'force_main_empty'
+          input_stream: 'force_loop_empty'
+          # Used to indicate "main" timestamp bound updates.
+          node {
+            calculator: 'GateCalculator'
+            input_stream: 'input'
+            input_stream: 'DISALLOW:force_main_empty'
+            output_stream: 'main'
+          }
+          node {
+            calculator: 'PreviousLoopbackCalculator'
+            input_stream: 'MAIN:main'
+            input_stream: 'LOOP:loop'
+            input_stream_info: { tag_index: 'LOOP' back_edge: true }
+            output_stream: 'PREV_LOOP:prev_loop'
+          }
+          node {
+            calculator: 'PassThroughCalculator'
+            input_stream: 'input'
+            input_stream: 'prev_loop'
+            output_stream: 'passed_through_input'
+            output_stream: 'passed_through_prev_loop'
+          }
+          # Used to indicate "loop" timestamp bound updates.
+          node {
+            calculator: 'GateCalculator'
+            input_stream: 'input'
+            input_stream: 'DISALLOW:force_loop_empty'
+            output_stream: 'loop'
+          }
+          node {
+            calculator: 'MakePairCalculator'
+            input_stream: 'passed_through_input'
+            input_stream: 'passed_through_prev_loop'
+            output_stream: 'passed_through_input_and_prev_loop'
+          }
+        )");
+    tool::AddVectorSink("passed_through_input_and_prev_loop", &graph_config,
+                        &output_packets_);
+    MP_ASSERT_OK(graph_.Initialize(graph_config, {}));
+    MP_ASSERT_OK(graph_.StartRun({}));
+  }
+
+  void SendPackets(int timestamp, int input, bool force_main_empty,
+                   bool force_loop_empty) {
+    MP_ASSERT_OK(graph_.AddPacketToInputStream(
+        "input", MakePacket<int>(input).At(Timestamp(timestamp))));
+    MP_ASSERT_OK(graph_.AddPacketToInputStream(
+        "force_main_empty",
+        MakePacket<bool>(force_main_empty).At(Timestamp(timestamp))));
+    MP_ASSERT_OK(graph_.AddPacketToInputStream(
+        "force_loop_empty",
+        MakePacket<bool>(force_loop_empty).At(Timestamp(timestamp))));
+  }
+
+  CalculatorGraph graph_;
+  std::vector<Packet> output_packets_;
+};
+
+TEST_F(PreviousLoopbackCalculatorProcessingTimestampsTest,
+       MultiplePacketsEmptyMainNonEmptyLoop) {
+  SendPackets(/*timestamp=*/1, /*input=*/1, /*force_main_empty=*/true,
+              /*force_loop_empty=*/false);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket()))));
+
+  SendPackets(/*timestamp=*/2, /*input=*/2, /*force_main_empty=*/true,
+              /*force_loop_empty=*/false);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())),
+                  PairPacket(Timestamp(2), Pair(IntPacket(2), EmptyPacket()))));
+
+  SendPackets(/*timestamp=*/3, /*input=*/3, /*force_main_empty=*/true,
+              /*force_loop_empty=*/false);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())),
+                  PairPacket(Timestamp(2), Pair(IntPacket(2), EmptyPacket())),
+                  PairPacket(Timestamp(3), Pair(IntPacket(3), EmptyPacket()))));
+
+  SendPackets(/*timestamp=*/5, /*input=*/5, /*force_main_empty=*/true,
+              /*force_loop_empty=*/false);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())),
+                  PairPacket(Timestamp(2), Pair(IntPacket(2), EmptyPacket())),
+                  PairPacket(Timestamp(3), Pair(IntPacket(3), EmptyPacket())),
+                  PairPacket(Timestamp(5), Pair(IntPacket(5), EmptyPacket()))));
+
+  SendPackets(/*timestamp=*/15, /*input=*/15,
+              /*force_main_empty=*/true,
+              /*force_loop_empty=*/false);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(
+          PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())),
+          PairPacket(Timestamp(2), Pair(IntPacket(2), EmptyPacket())),
+          PairPacket(Timestamp(3), Pair(IntPacket(3), EmptyPacket())),
+          PairPacket(Timestamp(5), Pair(IntPacket(5), EmptyPacket())),
+          PairPacket(Timestamp(15), Pair(IntPacket(15), EmptyPacket()))));
+
+  MP_EXPECT_OK(graph_.CloseAllInputStreams());
+  MP_EXPECT_OK(graph_.WaitUntilDone());
+}
+
+TEST_F(PreviousLoopbackCalculatorProcessingTimestampsTest,
+       MultiplePacketsNonEmptyMainEmptyLoop) {
+  SendPackets(/*timestamp=*/1, /*input=*/1,
+              /*force_main_empty=*/false,
+              /*force_loop_empty=*/true);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket()))));
+
+  SendPackets(/*timestamp=*/2, /*input=*/2,
+              /*force_main_empty=*/false,
+              /*force_loop_empty=*/true);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())),
+                  PairPacket(Timestamp(2), Pair(IntPacket(2), EmptyPacket()))));
+
+  SendPackets(/*timestamp=*/3, /*input=*/3,
+              /*force_main_empty=*/false,
+              /*force_loop_empty=*/true);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())),
+                  PairPacket(Timestamp(2), Pair(IntPacket(2), EmptyPacket())),
+                  PairPacket(Timestamp(3), Pair(IntPacket(3), EmptyPacket()))));
+
+  SendPackets(/*timestamp=*/5, /*input=*/5,
+              /*force_main_empty=*/false,
+              /*force_loop_empty=*/true);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())),
+                  PairPacket(Timestamp(2), Pair(IntPacket(2), EmptyPacket())),
+                  PairPacket(Timestamp(3), Pair(IntPacket(3), EmptyPacket())),
+                  PairPacket(Timestamp(5), Pair(IntPacket(5), EmptyPacket()))));
+
+  SendPackets(/*timestamp=*/15, /*input=*/15,
+              /*force_main_empty=*/false,
+              /*force_loop_empty=*/true);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(
+          PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())),
+          PairPacket(Timestamp(2), Pair(IntPacket(2), EmptyPacket())),
+          PairPacket(Timestamp(3), Pair(IntPacket(3), EmptyPacket())),
+          PairPacket(Timestamp(5), Pair(IntPacket(5), EmptyPacket())),
+          PairPacket(Timestamp(15), Pair(IntPacket(15), EmptyPacket()))));
+
+  MP_EXPECT_OK(graph_.CloseAllInputStreams());
+  MP_EXPECT_OK(graph_.WaitUntilDone());
+}
+
+TEST_F(PreviousLoopbackCalculatorProcessingTimestampsTest,
+       MultiplePacketsAlteringMainNonEmptyLoop) {
+  SendPackets(/*timestamp=*/1, /*input=*/1,
+              /*force_main_empty=*/false,
+              /*force_loop_empty=*/false);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket()))));
+
+  SendPackets(/*timestamp=*/2, /*input=*/2, /*force_main_empty=*/true,
+              /*force_loop_empty=*/false);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())),
+                  PairPacket(Timestamp(2), Pair(IntPacket(2), EmptyPacket()))));
+
+  SendPackets(/*timestamp=*/3, /*input=*/3,
+              /*force_main_empty=*/false,
+              /*force_loop_empty=*/false);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())),
+                  PairPacket(Timestamp(2), Pair(IntPacket(2), EmptyPacket())),
+                  PairPacket(Timestamp(3), Pair(IntPacket(3), IntPacket(1)))));
+
+  SendPackets(/*timestamp=*/5, /*input=*/5,
+              /*force_main_empty=*/false,
+              /*force_loop_empty=*/false);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())),
+                  PairPacket(Timestamp(2), Pair(IntPacket(2), EmptyPacket())),
+                  PairPacket(Timestamp(3), Pair(IntPacket(3), IntPacket(1))),
+                  PairPacket(Timestamp(5), Pair(IntPacket(5), IntPacket(3)))));
+
+  SendPackets(/*timestamp=*/15, /*input=*/15,
+              /*force_main_empty=*/true,
+              /*force_loop_empty=*/false);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(
+          PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())),
+          PairPacket(Timestamp(2), Pair(IntPacket(2), EmptyPacket())),
+          PairPacket(Timestamp(3), Pair(IntPacket(3), IntPacket(1))),
+          PairPacket(Timestamp(5), Pair(IntPacket(5), IntPacket(3))),
+          PairPacket(Timestamp(15), Pair(IntPacket(15), EmptyPacket()))));
+
+  MP_EXPECT_OK(graph_.CloseAllInputStreams());
+  MP_EXPECT_OK(graph_.WaitUntilDone());
+}
+
+TEST_F(PreviousLoopbackCalculatorProcessingTimestampsTest,
+       MultiplePacketsNonEmptyMainAlteringLoop) {
+  SendPackets(/*timestamp=*/1, /*input=*/1,
+              /*force_main_empty=*/false,
+              /*force_loop_empty=*/false);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket()))));
+
+  SendPackets(/*timestamp=*/2, /*input=*/2,
+              /*force_main_empty=*/false,
+              /*force_loop_empty=*/true);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())),
+                  PairPacket(Timestamp(2), Pair(IntPacket(2), IntPacket(1)))));
+
+  SendPackets(/*timestamp=*/3, /*input=*/3,
+              /*force_main_empty=*/false,
+              /*force_loop_empty=*/false);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())),
+                  PairPacket(Timestamp(2), Pair(IntPacket(2), IntPacket(1))),
+                  PairPacket(Timestamp(3), Pair(IntPacket(3), EmptyPacket()))));
+
+  SendPackets(/*timestamp=*/5, /*input=*/5,
+              /*force_main_empty=*/false,
+              /*force_loop_empty=*/true);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())),
+                  PairPacket(Timestamp(2), Pair(IntPacket(2), IntPacket(1))),
+                  PairPacket(Timestamp(3), Pair(IntPacket(3), EmptyPacket())),
+                  PairPacket(Timestamp(5), Pair(IntPacket(5), IntPacket(3)))));
+
+  SendPackets(/*timestamp=*/15, /*input=*/15,
+              /*force_main_empty=*/false,
+              /*force_loop_empty=*/false);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(
+          PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())),
+          PairPacket(Timestamp(2), Pair(IntPacket(2), IntPacket(1))),
+          PairPacket(Timestamp(3), Pair(IntPacket(3), EmptyPacket())),
+          PairPacket(Timestamp(5), Pair(IntPacket(5), IntPacket(3))),
+          PairPacket(Timestamp(15), Pair(IntPacket(15), EmptyPacket()))));
+
+  MP_EXPECT_OK(graph_.CloseAllInputStreams());
+  MP_EXPECT_OK(graph_.WaitUntilDone());
+}
+
+TEST_F(PreviousLoopbackCalculatorProcessingTimestampsTest,
+       MultiplePacketsCheckIfLastCorrectAlteringMainAlteringLoop) {
+  int num_packets = 1000;
+  for (int i = 0; i < num_packets; ++i) {
+    bool force_main_empty = i % 3 == 0 ? true : false;
+    bool force_loop_empty = i % 2 == 0 ? true : false;
+    SendPackets(/*timestamp=*/i + 1, /*input=*/i + 1, force_main_empty,
+                force_loop_empty);
+  }
+  SendPackets(/*timestamp=*/num_packets + 1,
+              /*input=*/num_packets + 1, /*force_main_empty=*/false,
+              /*force_loop_empty=*/false);
+  SendPackets(/*timestamp=*/num_packets + 2,
+              /*input=*/num_packets + 2, /*force_main_empty=*/false,
+              /*force_loop_empty=*/false);
+
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  ASSERT_FALSE(output_packets_.empty());
+  EXPECT_THAT(
+      output_packets_.back(),
+      PairPacket(Timestamp(num_packets + 2),
+                 Pair(IntPacket(num_packets + 2), IntPacket(num_packets + 1))));
+
+  MP_EXPECT_OK(graph_.CloseAllInputStreams());
+  MP_EXPECT_OK(graph_.WaitUntilDone());
+}
+
+// Similar to GateCalculator, but it doesn't propagate timestamp bound updates.
+class DroppingGateCalculator : public CalculatorBase {
+ public:
+  static ::mediapipe::Status GetContract(CalculatorContract* cc) {
+    cc->Inputs().Index(0).SetAny();
+    cc->Inputs().Tag("DISALLOW").Set<bool>();
+    cc->Outputs().Index(0).SetSameAs(&cc->Inputs().Index(0));
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Process(CalculatorContext* cc) final {
+    if (!cc->Inputs().Index(0).IsEmpty() &&
+        !cc->Inputs().Tag("DISALLOW").Get<bool>()) {
+      cc->Outputs().Index(0).AddPacket(cc->Inputs().Index(0).Value());
+    }
+    return ::mediapipe::OkStatus();
+  }
+};
+REGISTER_CALCULATOR(DroppingGateCalculator);
+
+// Tests PreviousLoopbackCalculator in cases when there are no "LOOP" timestamp
+// bound updates and non-empty packets for a while and the aforementioned start
+// to arrive at some point. So, "PREV_LOOP" is delayed for a couple of inputs.
+class PreviousLoopbackCalculatorDelayBehaviorTest : public testing::Test {
+ protected:
+  void SetUp() override {
+    CalculatorGraphConfig graph_config =
+        ParseTextProtoOrDie<CalculatorGraphConfig>(R"(
+          input_stream: 'input'
+          # Drops "loop" when set to "true", delaying output of prev_loop, hence
+          # delaying output of the graph.
+          input_stream: 'delay_next_output'
+          node {
+            calculator: 'PreviousLoopbackCalculator'
+            input_stream: 'MAIN:input'
+            input_stream: 'LOOP:loop'
+            input_stream_info: { tag_index: 'LOOP' back_edge: true }
+            output_stream: 'PREV_LOOP:prev_loop'
+          }
+          node {
+            calculator: 'PassThroughCalculator'
+            input_stream: 'input'
+            input_stream: 'prev_loop'
+            output_stream: 'passed_through_input'
+            output_stream: 'passed_through_prev_loop'
+          }
+          node {
+            calculator: 'DroppingGateCalculator'
+            input_stream: 'input'
+            input_stream: 'DISALLOW:delay_next_output'
+            output_stream: 'loop'
+          }
+          node {
+            calculator: 'MakePairCalculator'
+            input_stream: 'passed_through_input'
+            input_stream: 'passed_through_prev_loop'
+            output_stream: 'passed_through_input_and_prev_loop'
+          }
+        )");
+    tool::AddVectorSink("passed_through_input_and_prev_loop", &graph_config,
+                        &output_packets_);
+    MP_ASSERT_OK(graph_.Initialize(graph_config, {}));
+    MP_ASSERT_OK(graph_.StartRun({}));
+  }
+
+  void SendPackets(int timestamp, int input, bool delay_next_output) {
+    MP_ASSERT_OK(graph_.AddPacketToInputStream(
+        "input", MakePacket<int>(input).At(Timestamp(timestamp))));
+    MP_ASSERT_OK(graph_.AddPacketToInputStream(
+        "delay_next_output",
+        MakePacket<bool>(delay_next_output).At(Timestamp(timestamp))));
+  }
+
+  CalculatorGraph graph_;
+  std::vector<Packet> output_packets_;
+};
+
+TEST_F(PreviousLoopbackCalculatorDelayBehaviorTest, MultipleDelayedOutputs) {
+  SendPackets(/*timestamp=*/1, /*input=*/1, /*delay_next_output=*/true);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket()))));
+
+  SendPackets(/*timestamp=*/2, /*input=*/2, /*delay_next_output=*/true);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket()))));
+
+  SendPackets(/*timestamp=*/3, /*input=*/3, /*delay_next_output=*/true);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket()))));
+
+  SendPackets(/*timestamp=*/5, /*input=*/5, /*delay_next_output=*/false);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())),
+                  PairPacket(Timestamp(2), Pair(IntPacket(2), EmptyPacket())),
+                  PairPacket(Timestamp(3), Pair(IntPacket(3), EmptyPacket())),
+                  PairPacket(Timestamp(5), Pair(IntPacket(5), EmptyPacket()))));
+
+  SendPackets(/*timestamp=*/15, /*input=*/15, /*delay_next_output=*/false);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(
+          PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())),
+          PairPacket(Timestamp(2), Pair(IntPacket(2), EmptyPacket())),
+          PairPacket(Timestamp(3), Pair(IntPacket(3), EmptyPacket())),
+          PairPacket(Timestamp(5), Pair(IntPacket(5), EmptyPacket())),
+          PairPacket(Timestamp(15), Pair(IntPacket(15), IntPacket(5)))));
+
+  MP_EXPECT_OK(graph_.CloseAllInputStreams());
+  MP_EXPECT_OK(graph_.WaitUntilDone());
+}
+
+TEST_F(PreviousLoopbackCalculatorDelayBehaviorTest,
+       NonDelayedOutputFollowedByMultipleDelayedOutputs) {
+  SendPackets(/*timestamp=*/1, /*input=*/1, /*delay_next_output=*/false);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket()))));
+
+  SendPackets(/*timestamp=*/2, /*input=*/2, /*delay_next_output=*/true);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())),
+                  PairPacket(Timestamp(2), Pair(IntPacket(2), IntPacket(1)))));
+
+  SendPackets(/*timestamp=*/3, /*input=*/3, /*delay_next_output=*/true);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())),
+                  PairPacket(Timestamp(2), Pair(IntPacket(2), IntPacket(1)))));
+
+  SendPackets(/*timestamp=*/5, /*input=*/5, /*delay_next_output=*/false);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())),
+                  PairPacket(Timestamp(2), Pair(IntPacket(2), IntPacket(1))),
+                  PairPacket(Timestamp(3), Pair(IntPacket(3), EmptyPacket())),
+                  PairPacket(Timestamp(5), Pair(IntPacket(5), EmptyPacket()))));
+
+  SendPackets(/*timestamp=*/15, /*input=*/15, /*delay_next_output=*/false);
+  MP_EXPECT_OK(graph_.WaitUntilIdle());
+  EXPECT_THAT(
+      output_packets_,
+      ElementsAre(
+          PairPacket(Timestamp(1), Pair(IntPacket(1), EmptyPacket())),
+          PairPacket(Timestamp(2), Pair(IntPacket(2), IntPacket(1))),
+          PairPacket(Timestamp(3), Pair(IntPacket(3), EmptyPacket())),
+          PairPacket(Timestamp(5), Pair(IntPacket(5), EmptyPacket())),
+          PairPacket(Timestamp(15), Pair(IntPacket(15), IntPacket(5)))));
+
+  MP_EXPECT_OK(graph_.CloseAllInputStreams());
+  MP_EXPECT_OK(graph_.WaitUntilDone());
+}
+
 }  // anonymous namespace
 }  // namespace mediapipe
@@ -0,0 +1,165 @@
+// 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_SPLIT_NORMALIZED_LANDMARK_LIST_CALCULATOR_H_  // NOLINT
+#define MEDIAPIPE_CALCULATORS_CORE_SPLIT_NORMALIZED_LANDMARK_LIST_CALCULATOR_H_  // NOLINT
+
+#include "mediapipe/calculators/core/split_vector_calculator.pb.h"
+#include "mediapipe/framework/calculator_framework.h"
+#include "mediapipe/framework/formats/landmark.pb.h"
+#include "mediapipe/framework/port/canonical_errors.h"
+#include "mediapipe/framework/port/ret_check.h"
+#include "mediapipe/util/resource_util.h"
+
+namespace mediapipe {
+
+// Splits an input packet with NormalizedLandmarkList into
+// multiple NormalizedLandmarkList output packets using the [begin, end) ranges
+// specified in SplitVectorCalculatorOptions. If the option "element_only" is
+// set to true, all ranges should be of size 1 and all outputs will be elements
+// of type NormalizedLandmark. If "element_only" is false, ranges can be
+// non-zero in size and all outputs will be of type NormalizedLandmarkList.
+// If the option "combine_outputs" is set to true, only one output stream can be
+// specified and all ranges of elements will be combined into one
+// NormalizedLandmarkList.
+class SplitNormalizedLandmarkListCalculator : public CalculatorBase {
+ public:
+  static ::mediapipe::Status GetContract(CalculatorContract* cc) {
+    RET_CHECK(cc->Inputs().NumEntries() == 1);
+    RET_CHECK(cc->Outputs().NumEntries() != 0);
+
+    cc->Inputs().Index(0).Set<NormalizedLandmarkList>();
+
+    const auto& options =
+        cc->Options<::mediapipe::SplitVectorCalculatorOptions>();
+
+    if (options.combine_outputs()) {
+      RET_CHECK_EQ(cc->Outputs().NumEntries(), 1);
+      cc->Outputs().Index(0).Set<NormalizedLandmarkList>();
+      for (int i = 0; i < options.ranges_size() - 1; ++i) {
+        for (int j = i + 1; j < options.ranges_size(); ++j) {
+          const auto& range_0 = options.ranges(i);
+          const auto& range_1 = options.ranges(j);
+          if ((range_0.begin() >= range_1.begin() &&
+               range_0.begin() < range_1.end()) ||
+              (range_1.begin() >= range_0.begin() &&
+               range_1.begin() < range_0.end())) {
+            return ::mediapipe::InvalidArgumentError(
+                "Ranges must be non-overlapping when using combine_outputs "
+                "option.");
+          }
+        }
+      }
+    } else {
+      if (cc->Outputs().NumEntries() != options.ranges_size()) {
+        return ::mediapipe::InvalidArgumentError(
+            "The number of output streams should match the number of ranges "
+            "specified in the CalculatorOptions.");
+      }
+
+      // Set the output types for each output stream.
+      for (int i = 0; i < cc->Outputs().NumEntries(); ++i) {
+        if (options.ranges(i).begin() < 0 || options.ranges(i).end() < 0 ||
+            options.ranges(i).begin() >= options.ranges(i).end()) {
+          return ::mediapipe::InvalidArgumentError(
+              "Indices should be non-negative and begin index should be less "
+              "than the end index.");
+        }
+        if (options.element_only()) {
+          if (options.ranges(i).end() - options.ranges(i).begin() != 1) {
+            return ::mediapipe::InvalidArgumentError(
+                "Since element_only is true, all ranges should be of size 1.");
+          }
+          cc->Outputs().Index(i).Set<NormalizedLandmark>();
+        } else {
+          cc->Outputs().Index(i).Set<NormalizedLandmarkList>();
+        }
+      }
+    }
+
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Open(CalculatorContext* cc) override {
+    cc->SetOffset(TimestampDiff(0));
+
+    const auto& options =
+        cc->Options<::mediapipe::SplitVectorCalculatorOptions>();
+
+    element_only_ = options.element_only();
+    combine_outputs_ = options.combine_outputs();
+
+    for (const auto& range : options.ranges()) {
+      ranges_.push_back({range.begin(), range.end()});
+      max_range_end_ = std::max(max_range_end_, range.end());
+      total_elements_ += range.end() - range.begin();
+    }
+
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Process(CalculatorContext* cc) override {
+    const NormalizedLandmarkList& input =
+        cc->Inputs().Index(0).Get<NormalizedLandmarkList>();
+    RET_CHECK_GE(input.landmark_size(), max_range_end_);
+
+    if (combine_outputs_) {
+      NormalizedLandmarkList output;
+      for (int i = 0; i < ranges_.size(); ++i) {
+        for (int j = ranges_[i].first; j < ranges_[i].second; ++j) {
+          const NormalizedLandmark& input_landmark = input.landmark(j);
+          *output.add_landmark() = input_landmark;
+        }
+      }
+      RET_CHECK_EQ(output.landmark_size(), total_elements_);
+      cc->Outputs().Index(0).AddPacket(
+          MakePacket<NormalizedLandmarkList>(output).At(cc->InputTimestamp()));
+    } else {
+      if (element_only_) {
+        for (int i = 0; i < ranges_.size(); ++i) {
+          cc->Outputs().Index(i).AddPacket(
+              MakePacket<NormalizedLandmark>(input.landmark(ranges_[i].first))
+                  .At(cc->InputTimestamp()));
+        }
+      } else {
+        for (int i = 0; i < ranges_.size(); ++i) {
+          NormalizedLandmarkList output;
+          for (int j = ranges_[i].first; j < ranges_[i].second; ++j) {
+            const NormalizedLandmark& input_landmark = input.landmark(j);
+            *output.add_landmark() = input_landmark;
+          }
+          cc->Outputs().Index(i).AddPacket(
+              MakePacket<NormalizedLandmarkList>(output).At(
+                  cc->InputTimestamp()));
+        }
+      }
+    }
+
+    return ::mediapipe::OkStatus();
+  }
+
+ private:
+  std::vector<std::pair<int32, int32>> ranges_;
+  int32 max_range_end_ = -1;
+  int32 total_elements_ = 0;
+  bool element_only_ = false;
+  bool combine_outputs_ = false;
+};
+
+REGISTER_CALCULATOR(SplitNormalizedLandmarkListCalculator);
+
+}  // namespace mediapipe
+
+// NOLINTNEXTLINE
+#endif  // MEDIAPIPE_CALCULATORS_CORE_SPLIT_NORMALIZED_LANDMARK_LIST_CALCULATOR_H_
@@ -0,0 +1,404 @@
+// 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 <vector>
+
+#include "mediapipe/calculators/core/split_vector_calculator.pb.h"
+#include "mediapipe/framework/calculator_framework.h"
+#include "mediapipe/framework/calculator_runner.h"
+#include "mediapipe/framework/deps/file_path.h"
+#include "mediapipe/framework/formats/landmark.pb.h"
+#include "mediapipe/framework/port/gmock.h"
+#include "mediapipe/framework/port/gtest.h"
+#include "mediapipe/framework/port/integral_types.h"
+#include "mediapipe/framework/port/parse_text_proto.h"
+#include "mediapipe/framework/port/status_matchers.h"  // NOLINT
+#include "mediapipe/framework/tool/validate_type.h"
+
+namespace mediapipe {
+
+constexpr float kLocationVal = 3;
+
+class SplitNormalizedLandmarkListCalculatorTest : public ::testing::Test {
+ protected:
+  void TearDown() { expected_landmarks_.reset(); }
+
+  void PrepareNormalizedLandmarkList(int list_size) {
+    // Prepare input landmark list.
+    input_landmarks_ = absl::make_unique<NormalizedLandmarkList>();
+    expected_landmarks_ = absl::make_unique<NormalizedLandmarkList>();
+    for (int i = 0; i < list_size; ++i) {
+      NormalizedLandmark* landmark = input_landmarks_->add_landmark();
+      landmark->set_x(i * kLocationVal);
+      landmark->set_y(i * kLocationVal);
+      landmark->set_z(i * kLocationVal);
+      // Save the landmarks for comparison after the graph runs.
+      *expected_landmarks_->add_landmark() = *landmark;
+    }
+  }
+
+  void ValidateListOutput(std::vector<Packet>& output_packets,
+                          int expected_elements, int input_begin_index) {
+    ASSERT_EQ(1, output_packets.size());
+    const NormalizedLandmarkList& output_landmarks =
+        output_packets[0].Get<NormalizedLandmarkList>();
+    ASSERT_EQ(expected_elements, output_landmarks.landmark_size());
+
+    for (int i = 0; i < expected_elements; ++i) {
+      const NormalizedLandmark& expected_landmark =
+          expected_landmarks_->landmark(input_begin_index + i);
+      const NormalizedLandmark& result = output_landmarks.landmark(i);
+      EXPECT_FLOAT_EQ(expected_landmark.x(), result.x());
+      EXPECT_FLOAT_EQ(expected_landmark.y(), result.y());
+      EXPECT_FLOAT_EQ(expected_landmark.z(), result.z());
+    }
+  }
+
+  void ValidateCombinedListOutput(std::vector<Packet>& output_packets,
+                                  int expected_elements,
+                                  std::vector<int>& input_begin_indices,
+                                  std::vector<int>& input_end_indices) {
+    ASSERT_EQ(1, output_packets.size());
+    ASSERT_EQ(input_begin_indices.size(), input_end_indices.size());
+    const NormalizedLandmarkList& output_landmarks =
+        output_packets[0].Get<NormalizedLandmarkList>();
+    ASSERT_EQ(expected_elements, output_landmarks.landmark_size());
+    const int num_ranges = input_begin_indices.size();
+
+    int element_id = 0;
+    for (int range_id = 0; range_id < num_ranges; ++range_id) {
+      for (int i = input_begin_indices[range_id];
+           i < input_end_indices[range_id]; ++i) {
+        const NormalizedLandmark& expected_landmark =
+            expected_landmarks_->landmark(i);
+        const NormalizedLandmark& result =
+            output_landmarks.landmark(element_id);
+        EXPECT_FLOAT_EQ(expected_landmark.x(), result.x());
+        EXPECT_FLOAT_EQ(expected_landmark.y(), result.y());
+        EXPECT_FLOAT_EQ(expected_landmark.z(), result.z());
+        element_id++;
+      }
+    }
+  }
+
+  void ValidateElementOutput(std::vector<Packet>& output_packets,
+                             int input_begin_index) {
+    ASSERT_EQ(1, output_packets.size());
+
+    const NormalizedLandmark& output_landmark =
+        output_packets[0].Get<NormalizedLandmark>();
+    ASSERT_TRUE(output_landmark.IsInitialized());
+
+    const NormalizedLandmark& expected_landmark =
+        expected_landmarks_->landmark(input_begin_index);
+
+    EXPECT_FLOAT_EQ(expected_landmark.x(), output_landmark.x());
+    EXPECT_FLOAT_EQ(expected_landmark.y(), output_landmark.y());
+    EXPECT_FLOAT_EQ(expected_landmark.z(), output_landmark.z());
+  }
+
+  std::unique_ptr<NormalizedLandmarkList> input_landmarks_ = nullptr;
+  std::unique_ptr<NormalizedLandmarkList> expected_landmarks_ = nullptr;
+  std::unique_ptr<CalculatorRunner> runner_ = nullptr;
+};
+
+TEST_F(SplitNormalizedLandmarkListCalculatorTest, SmokeTest) {
+  PrepareNormalizedLandmarkList(/*list_size=*/5);
+  ASSERT_NE(input_landmarks_, nullptr);
+
+  // Prepare a graph to use the SplitNormalizedLandmarkListCalculator.
+  CalculatorGraphConfig graph_config =
+      ::mediapipe::ParseTextProtoOrDie<CalculatorGraphConfig>(
+          R"(
+            input_stream: "landmarks_in"
+            node {
+              calculator: "SplitNormalizedLandmarkListCalculator"
+              input_stream: "landmarks_in"
+              output_stream: "range_0"
+              output_stream: "range_1"
+              output_stream: "range_2"
+              options {
+                [mediapipe.SplitVectorCalculatorOptions.ext] {
+                  ranges: { begin: 0 end: 1 }
+                  ranges: { begin: 1 end: 4 }
+                  ranges: { begin: 4 end: 5 }
+                }
+              }
+            }
+          )");
+  std::vector<Packet> range_0_packets;
+  tool::AddVectorSink("range_0", &graph_config, &range_0_packets);
+  std::vector<Packet> range_1_packets;
+  tool::AddVectorSink("range_1", &graph_config, &range_1_packets);
+  std::vector<Packet> range_2_packets;
+  tool::AddVectorSink("range_2", &graph_config, &range_2_packets);
+
+  // Run the graph.
+  CalculatorGraph graph;
+  MP_ASSERT_OK(graph.Initialize(graph_config));
+  MP_ASSERT_OK(graph.StartRun({}));
+  MP_ASSERT_OK(graph.AddPacketToInputStream(
+      "landmarks_in", Adopt(input_landmarks_.release()).At(Timestamp(0))));
+  // Wait until the calculator finishes processing.
+  MP_ASSERT_OK(graph.WaitUntilIdle());
+
+  ValidateListOutput(range_0_packets, /*expected_elements=*/1,
+                     /*input_begin_index=*/0);
+  ValidateListOutput(range_1_packets, /*expected_elements=*/3,
+                     /*input_begin_index=*/1);
+  ValidateListOutput(range_2_packets, /*expected_elements=*/1,
+                     /*input_begin_index=*/4);
+
+  // Fully close the graph at the end.
+  MP_ASSERT_OK(graph.CloseInputStream("landmarks_in"));
+  MP_ASSERT_OK(graph.WaitUntilDone());
+}
+
+TEST_F(SplitNormalizedLandmarkListCalculatorTest, InvalidRangeTest) {
+  // Prepare a graph to use the SplitNormalizedLandmarkListCalculator.
+  CalculatorGraphConfig graph_config =
+      ::mediapipe::ParseTextProtoOrDie<CalculatorGraphConfig>(
+          R"(
+            input_stream: "landmarks_in"
+            node {
+              calculator: "SplitNormalizedLandmarkListCalculator"
+              input_stream: "landmarks_in"
+              output_stream: "range_0"
+              options {
+                [mediapipe.SplitVectorCalculatorOptions.ext] {
+                  ranges: { begin: 0 end: 0 }
+                }
+              }
+            }
+          )");
+
+  // Run the graph.
+  CalculatorGraph graph;
+  // The graph should fail running because of an invalid range (begin == end).
+  ASSERT_FALSE(graph.Initialize(graph_config).ok());
+}
+
+TEST_F(SplitNormalizedLandmarkListCalculatorTest,
+       InvalidOutputStreamCountTest) {
+  // Prepare a graph to use the SplitNormalizedLandmarkListCalculator.
+  CalculatorGraphConfig graph_config =
+      ::mediapipe::ParseTextProtoOrDie<CalculatorGraphConfig>(
+          R"(
+            input_stream: "landmarks_in"
+            node {
+              calculator: "SplitNormalizedLandmarkListCalculator"
+              input_stream: "landmarks_in"
+              output_stream: "range_0"
+              output_stream: "range_1"
+              options {
+                [mediapipe.SplitVectorCalculatorOptions.ext] {
+                  ranges: { begin: 0 end: 1 }
+                }
+              }
+            }
+          )");
+
+  // Run the graph.
+  CalculatorGraph graph;
+  // The graph should fail running because the number of output streams does not
+  // match the number of range elements in the options.
+  ASSERT_FALSE(graph.Initialize(graph_config).ok());
+}
+
+TEST_F(SplitNormalizedLandmarkListCalculatorTest,
+       InvalidCombineOutputsMultipleOutputsTest) {
+  // Prepare a graph to use the SplitNormalizedLandmarkListCalculator.
+  CalculatorGraphConfig graph_config =
+      ::mediapipe::ParseTextProtoOrDie<CalculatorGraphConfig>(
+          R"(
+            input_stream: "landmarks_in"
+            node {
+              calculator: "SplitNormalizedLandmarkListCalculator"
+              input_stream: "landmarks_in"
+              output_stream: "range_0"
+              output_stream: "range_1"
+              options {
+                [mediapipe.SplitVectorCalculatorOptions.ext] {
+                  ranges: { begin: 0 end: 1 }
+                  ranges: { begin: 2 end: 3 }
+                  combine_outputs: true
+                }
+              }
+            }
+          )");
+
+  // Run the graph.
+  CalculatorGraph graph;
+  // The graph should fail running because the number of output streams does not
+  // match the number of range elements in the options.
+  ASSERT_FALSE(graph.Initialize(graph_config).ok());
+}
+
+TEST_F(SplitNormalizedLandmarkListCalculatorTest,
+       InvalidOverlappingRangesTest) {
+  // Prepare a graph to use the SplitNormalizedLandmarkListCalculator.
+  CalculatorGraphConfig graph_config =
+      ::mediapipe::ParseTextProtoOrDie<CalculatorGraphConfig>(
+          R"(
+            input_stream: "landmarks_in"
+            node {
+              calculator: "SplitNormalizedLandmarkListCalculator"
+              input_stream: "landmarks_in"
+              output_stream: "range_0"
+              options {
+                [mediapipe.SplitVectorCalculatorOptions.ext] {
+                  ranges: { begin: 0 end: 3 }
+                  ranges: { begin: 1 end: 4 }
+                  combine_outputs: true
+                }
+              }
+            }
+          )");
+
+  // Run the graph.
+  CalculatorGraph graph;
+  // The graph should fail running because there are overlapping ranges.
+  ASSERT_FALSE(graph.Initialize(graph_config).ok());
+}
+
+TEST_F(SplitNormalizedLandmarkListCalculatorTest, SmokeTestElementOnly) {
+  PrepareNormalizedLandmarkList(/*list_size=*/5);
+  ASSERT_NE(input_landmarks_, nullptr);
+
+  // Prepare a graph to use the SplitNormalizedLandmarkListCalculator.
+  CalculatorGraphConfig graph_config =
+      ::mediapipe::ParseTextProtoOrDie<CalculatorGraphConfig>(
+          R"(
+            input_stream: "landmarks_in"
+            node {
+              calculator: "SplitNormalizedLandmarkListCalculator"
+              input_stream: "landmarks_in"
+              output_stream: "range_0"
+              output_stream: "range_1"
+              output_stream: "range_2"
+              options {
+                [mediapipe.SplitVectorCalculatorOptions.ext] {
+                  ranges: { begin: 0 end: 1 }
+                  ranges: { begin: 2 end: 3 }
+                  ranges: { begin: 4 end: 5 }
+                  element_only: true
+                }
+              }
+            }
+          )");
+  std::vector<Packet> range_0_packets;
+  tool::AddVectorSink("range_0", &graph_config, &range_0_packets);
+  std::vector<Packet> range_1_packets;
+  tool::AddVectorSink("range_1", &graph_config, &range_1_packets);
+  std::vector<Packet> range_2_packets;
+  tool::AddVectorSink("range_2", &graph_config, &range_2_packets);
+
+  // Run the graph.
+  CalculatorGraph graph;
+  MP_ASSERT_OK(graph.Initialize(graph_config));
+  MP_ASSERT_OK(graph.StartRun({}));
+  MP_ASSERT_OK(graph.AddPacketToInputStream(
+      "landmarks_in", Adopt(input_landmarks_.release()).At(Timestamp(0))));
+  // Wait until the calculator finishes processing.
+  MP_ASSERT_OK(graph.WaitUntilIdle());
+
+  ValidateElementOutput(range_0_packets,
+                        /*input_begin_index=*/0);
+  ValidateElementOutput(range_1_packets,
+                        /*input_begin_index=*/2);
+  ValidateElementOutput(range_2_packets,
+                        /*input_begin_index=*/4);
+
+  // Fully close the graph at the end.
+  MP_ASSERT_OK(graph.CloseInputStream("landmarks_in"));
+  MP_ASSERT_OK(graph.WaitUntilDone());
+}
+
+TEST_F(SplitNormalizedLandmarkListCalculatorTest, SmokeTestCombiningOutputs) {
+  PrepareNormalizedLandmarkList(/*list_size=*/5);
+  ASSERT_NE(input_landmarks_, nullptr);
+
+  // Prepare a graph to use the SplitNormalizedLandmarkListCalculator.
+  CalculatorGraphConfig graph_config =
+      ::mediapipe::ParseTextProtoOrDie<CalculatorGraphConfig>(
+          R"(
+            input_stream: "landmarks_in"
+            node {
+              calculator: "SplitNormalizedLandmarkListCalculator"
+              input_stream: "landmarks_in"
+              output_stream: "range_0"
+              options {
+                [mediapipe.SplitVectorCalculatorOptions.ext] {
+                  ranges: { begin: 0 end: 1 }
+                  ranges: { begin: 2 end: 3 }
+                  ranges: { begin: 4 end: 5 }
+                  combine_outputs: true
+                }
+              }
+            }
+          )");
+  std::vector<Packet> range_0_packets;
+  tool::AddVectorSink("range_0", &graph_config, &range_0_packets);
+
+  // Run the graph.
+  CalculatorGraph graph;
+  MP_ASSERT_OK(graph.Initialize(graph_config));
+  MP_ASSERT_OK(graph.StartRun({}));
+  MP_ASSERT_OK(graph.AddPacketToInputStream(
+      "landmarks_in", Adopt(input_landmarks_.release()).At(Timestamp(0))));
+  // Wait until the calculator finishes processing.
+  MP_ASSERT_OK(graph.WaitUntilIdle());
+
+  std::vector<int> input_begin_indices = {0, 2, 4};
+  std::vector<int> input_end_indices = {1, 3, 5};
+  ValidateCombinedListOutput(range_0_packets, /*expected_elements=*/3,
+                             input_begin_indices, input_end_indices);
+
+  // Fully close the graph at the end.
+  MP_ASSERT_OK(graph.CloseInputStream("landmarks_in"));
+  MP_ASSERT_OK(graph.WaitUntilDone());
+}
+
+TEST_F(SplitNormalizedLandmarkListCalculatorTest,
+       ElementOnlyDisablesVectorOutputs) {
+  // Prepare a graph to use the SplitNormalizedLandmarkListCalculator.
+  CalculatorGraphConfig graph_config =
+      ::mediapipe::ParseTextProtoOrDie<CalculatorGraphConfig>(
+          R"(
+            input_stream: "landmarks_in"
+            node {
+              calculator: "SplitNormalizedLandmarkListCalculator"
+              input_stream: "landmarks_in"
+              output_stream: "range_0"
+              output_stream: "range_1"
+              output_stream: "range_2"
+              options {
+                [mediapipe.SplitVectorCalculatorOptions.ext] {
+                  ranges: { begin: 0 end: 1 }
+                  ranges: { begin: 1 end: 4 }
+                  ranges: { begin: 4 end: 5 }
+                  element_only: true
+                }
+              }
+            }
+          )");
+
+  // Run the graph.
+  CalculatorGraph graph;
+  ASSERT_FALSE(graph.Initialize(graph_config).ok());
+}
+
+}  // namespace mediapipe
@@ -16,7 +16,9 @@

 #include <vector>

+#include "mediapipe/framework/formats/detection.pb.h"
 #include "mediapipe/framework/formats/landmark.pb.h"
+#include "mediapipe/framework/formats/matrix.h"
 #include "mediapipe/framework/formats/rect.pb.h"
 #include "tensorflow/lite/interpreter.h"

@@ -48,13 +50,25 @@ typedef SplitVectorCalculator<::mediapipe::NormalizedLandmark, false>
    SplitLandmarkVectorCalculator;
 REGISTER_CALCULATOR(SplitLandmarkVectorCalculator);

+typedef SplitVectorCalculator<::mediapipe::NormalizedLandmarkList, false>
+    SplitNormalizedLandmarkListVectorCalculator;
+REGISTER_CALCULATOR(SplitNormalizedLandmarkListVectorCalculator);
+
 typedef SplitVectorCalculator<::mediapipe::NormalizedRect, false>
    SplitNormalizedRectVectorCalculator;
 REGISTER_CALCULATOR(SplitNormalizedRectVectorCalculator);

+typedef SplitVectorCalculator<Matrix, false> SplitMatrixVectorCalculator;
+REGISTER_CALCULATOR(SplitMatrixVectorCalculator);
+
 #if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
 typedef SplitVectorCalculator<::tflite::gpu::gl::GlBuffer, true>
    MovableSplitGlBufferVectorCalculator;
 REGISTER_CALCULATOR(MovableSplitGlBufferVectorCalculator);
 #endif
+
+typedef SplitVectorCalculator<::mediapipe::Detection, false>
+    SplitDetectionVectorCalculator;
+REGISTER_CALCULATOR(SplitDetectionVectorCalculator);
+
 }  // namespace mediapipe
@@ -86,6 +86,14 @@ mediapipe_cc_proto_library(
    deps = [":opencv_image_encoder_calculator_proto"],
 )

+mediapipe_cc_proto_library(
+    name = "opencv_encoded_image_to_image_frame_calculator_cc_proto",
+    srcs = ["opencv_encoded_image_to_image_frame_calculator.proto"],
+    cc_deps = ["//mediapipe/framework:calculator_cc_proto"],
+    visibility = ["//visibility:public"],
+    deps = [":opencv_encoded_image_to_image_frame_calculator_proto"],
+)
+
 mediapipe_cc_proto_library(
    name = "mask_overlay_calculator_cc_proto",
    srcs = ["mask_overlay_calculator.proto"],
@@ -172,6 +180,7 @@ cc_library(
    srcs = ["opencv_encoded_image_to_image_frame_calculator.cc"],
    visibility = ["//visibility:public"],
    deps = [
+        ":opencv_encoded_image_to_image_frame_calculator_cc_proto",
        "//mediapipe/framework:calculator_framework",
        "//mediapipe/framework/formats:image_frame_opencv",
        "//mediapipe/framework/port:opencv_imgcodecs",
@@ -422,9 +431,12 @@ cc_library(
        ":recolor_calculator_cc_proto",
        "//mediapipe/framework:calculator_framework",
        "//mediapipe/framework/formats:image_frame",
+        "//mediapipe/framework/formats:image_frame_opencv",
        "//mediapipe/framework/port:status",
        "//mediapipe/framework/port:ret_check",
        "//mediapipe/util:color_cc_proto",
+        "//mediapipe/framework/port:opencv_core",
+        "//mediapipe/framework/port:opencv_imgproc",
    ] + select({
        "//mediapipe/gpu:disable_gpu": [],
        "//conditions:default": [
@@ -554,6 +566,27 @@ proto_library(
    deps = ["//mediapipe/framework:calculator_proto"],
 )

+proto_library(
+    name = "opencv_encoded_image_to_image_frame_calculator_proto",
+    srcs = ["opencv_encoded_image_to_image_frame_calculator.proto"],
+    visibility = ["//visibility:public"],
+    deps = ["//mediapipe/framework:calculator_proto"],
+)
+
+proto_library(
+    name = "feature_detector_calculator_proto",
+    srcs = ["feature_detector_calculator.proto"],
+    deps = ["//mediapipe/framework:calculator_proto"],
+)
+
+mediapipe_cc_proto_library(
+    name = "feature_detector_calculator_cc_proto",
+    srcs = ["feature_detector_calculator.proto"],
+    cc_deps = ["//mediapipe/framework:calculator_cc_proto"],
+    visibility = ["//visibility:public"],
+    deps = [":feature_detector_calculator_proto"],
+)
+
 cc_library(
    name = "mask_overlay_calculator",
    srcs = ["mask_overlay_calculator.cc"],
@@ -569,3 +602,30 @@ cc_library(
    ],
    alwayslink = 1,
 )
+
+cc_library(
+    name = "feature_detector_calculator",
+    srcs = ["feature_detector_calculator.cc"],
+    visibility = ["//mediapipe:__subpackages__"],
+    deps = [
+        ":feature_detector_calculator_cc_proto",
+        "//mediapipe/framework:calculator_framework",
+        "//mediapipe/framework/formats:image_frame",
+        "//mediapipe/framework/formats:image_frame_opencv",
+        "//mediapipe/framework/formats:landmark_cc_proto",
+        "//mediapipe/framework/formats:video_stream_header",
+        "//mediapipe/framework/port:integral_types",
+        "//mediapipe/framework/port:logging",
+        "//mediapipe/framework/port:opencv_core",
+        "//mediapipe/framework/port:opencv_features2d",
+        "//mediapipe/framework/port:opencv_imgproc",
+        "//mediapipe/framework/port:ret_check",
+        "//mediapipe/framework/port:status",
+        "//mediapipe/framework/port:threadpool",
+        "//mediapipe/framework/tool:options_util",
+        "@com_google_absl//absl/memory",
+        "@com_google_absl//absl/synchronization",
+        "@org_tensorflow//tensorflow/lite:framework",
+    ],
+    alwayslink = 1,
+)
@@ -0,0 +1,210 @@
+// Copyright 2020 The MediaPipe Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <memory>
+#include <vector>
+
+#include "absl/memory/memory.h"
+#include "absl/synchronization/blocking_counter.h"
+#include "mediapipe/calculators/image/feature_detector_calculator.pb.h"
+#include "mediapipe/framework/calculator_framework.h"
+#include "mediapipe/framework/formats/image_frame.h"
+#include "mediapipe/framework/formats/image_frame_opencv.h"
+#include "mediapipe/framework/formats/landmark.pb.h"
+#include "mediapipe/framework/formats/video_stream_header.h"
+#include "mediapipe/framework/port/integral_types.h"
+#include "mediapipe/framework/port/logging.h"
+#include "mediapipe/framework/port/opencv_core_inc.h"
+#include "mediapipe/framework/port/opencv_features2d_inc.h"
+#include "mediapipe/framework/port/opencv_imgproc_inc.h"
+#include "mediapipe/framework/port/ret_check.h"
+#include "mediapipe/framework/port/status.h"
+#include "mediapipe/framework/port/threadpool.h"
+#include "mediapipe/framework/tool/options_util.h"
+#include "tensorflow/lite/interpreter.h"
+
+namespace mediapipe {
+
+const char kOptionsTag[] = "OPTIONS";
+const int kPatchSize = 32;
+const int kNumThreads = 16;
+
+// A calculator to apply local feature detection.
+// Input stream:
+//   IMAGE: Input image frame of type ImageFrame from video stream.
+// Output streams:
+//   FEATURES: The detected keypoints from input image as vector<cv::KeyPoint>.
+//   PATCHES:  Optional output the extracted patches as vector<cv::Mat>
+class FeatureDetectorCalculator : public CalculatorBase {
+ public:
+  ~FeatureDetectorCalculator() override = default;
+
+  static ::mediapipe::Status GetContract(CalculatorContract* cc);
+
+  ::mediapipe::Status Open(CalculatorContext* cc) override;
+  ::mediapipe::Status Process(CalculatorContext* cc) override;
+
+ private:
+  FeatureDetectorCalculatorOptions options_;
+  cv::Ptr<cv::Feature2D> feature_detector_;
+  std::unique_ptr<::mediapipe::ThreadPool> pool_;
+
+  // Create image pyramid based on input image.
+  void ComputeImagePyramid(const cv::Mat& input_image,
+                           std::vector<cv::Mat>* image_pyramid);
+
+  // Extract the patch for single feature with image pyramid.
+  cv::Mat ExtractPatch(const cv::KeyPoint& feature,
+                       const std::vector<cv::Mat>& image_pyramid);
+};
+
+REGISTER_CALCULATOR(FeatureDetectorCalculator);
+
+::mediapipe::Status FeatureDetectorCalculator::GetContract(
+    CalculatorContract* cc) {
+  if (cc->Inputs().HasTag("IMAGE")) {
+    cc->Inputs().Tag("IMAGE").Set<ImageFrame>();
+  }
+  if (cc->Outputs().HasTag("FEATURES")) {
+    cc->Outputs().Tag("FEATURES").Set<std::vector<cv::KeyPoint>>();
+  }
+  if (cc->Outputs().HasTag("LANDMARKS")) {
+    cc->Outputs().Tag("LANDMARKS").Set<NormalizedLandmarkList>();
+  }
+  if (cc->Outputs().HasTag("PATCHES")) {
+    cc->Outputs().Tag("PATCHES").Set<std::vector<TfLiteTensor>>();
+  }
+  return ::mediapipe::OkStatus();
+}
+
+::mediapipe::Status FeatureDetectorCalculator::Open(CalculatorContext* cc) {
+  options_ =
+      tool::RetrieveOptions(cc->Options(), cc->InputSidePackets(), kOptionsTag)
+          .GetExtension(FeatureDetectorCalculatorOptions::ext);
+  feature_detector_ = cv::ORB::create(
+      options_.max_features(), options_.scale_factor(),
+      options_.pyramid_level(), kPatchSize - 1, 0, 2, cv::ORB::FAST_SCORE);
+  pool_ = absl::make_unique<::mediapipe::ThreadPool>("ThreadPool", kNumThreads);
+  pool_->StartWorkers();
+  return ::mediapipe::OkStatus();
+}
+
+::mediapipe::Status FeatureDetectorCalculator::Process(CalculatorContext* cc) {
+  const Timestamp& timestamp = cc->InputTimestamp();
+  if (timestamp == Timestamp::PreStream()) {
+    // Indicator packet.
+    return ::mediapipe::OkStatus();
+  }
+  InputStream* input_frame = &(cc->Inputs().Tag("IMAGE"));
+  cv::Mat input_view = formats::MatView(&input_frame->Get<ImageFrame>());
+  cv::Mat grayscale_view;
+  cv::cvtColor(input_view, grayscale_view, cv::COLOR_RGB2GRAY);
+
+  std::vector<cv::KeyPoint> keypoints;
+  feature_detector_->detect(grayscale_view, keypoints);
+  if (keypoints.size() > options_.max_features()) {
+    keypoints.resize(options_.max_features());
+  }
+
+  if (cc->Outputs().HasTag("FEATURES")) {
+    auto features_ptr = absl::make_unique<std::vector<cv::KeyPoint>>(keypoints);
+    cc->Outputs().Tag("FEATURES").Add(features_ptr.release(), timestamp);
+  }
+
+  if (cc->Outputs().HasTag("LANDMARKS")) {
+    auto landmarks_ptr = absl::make_unique<NormalizedLandmarkList>();
+    for (int j = 0; j < keypoints.size(); ++j) {
+      auto feature_landmark = landmarks_ptr->add_landmark();
+      feature_landmark->set_x(keypoints[j].pt.x / grayscale_view.cols);
+      feature_landmark->set_y(keypoints[j].pt.y / grayscale_view.rows);
+    }
+    cc->Outputs().Tag("LANDMARKS").Add(landmarks_ptr.release(), timestamp);
+  }
+
+  if (cc->Outputs().HasTag("PATCHES")) {
+    std::vector<cv::Mat> image_pyramid;
+    ComputeImagePyramid(grayscale_view, &image_pyramid);
+    std::vector<cv::Mat> patch_mat;
+    patch_mat.resize(keypoints.size());
+    absl::BlockingCounter counter(keypoints.size());
+    for (int i = 0; i < keypoints.size(); i++) {
+      pool_->Schedule(
+          [this, &image_pyramid, &keypoints, &patch_mat, i, &counter] {
+            patch_mat[i] = ExtractPatch(keypoints[i], image_pyramid);
+            counter.DecrementCount();
+          });
+    }
+    counter.Wait();
+    const int batch_size = options_.max_features();
+    auto patches = absl::make_unique<std::vector<TfLiteTensor>>();
+    TfLiteTensor tensor;
+    tensor.type = kTfLiteFloat32;
+    tensor.dims = TfLiteIntArrayCreate(4);
+    tensor.dims->data[0] = batch_size;
+    tensor.dims->data[1] = kPatchSize;
+    tensor.dims->data[2] = kPatchSize;
+    tensor.dims->data[3] = 1;
+    int num_bytes = batch_size * kPatchSize * kPatchSize * sizeof(float);
+    tensor.data.data = malloc(num_bytes);
+    tensor.bytes = num_bytes;
+    tensor.allocation_type = kTfLiteArenaRw;
+    float* tensor_buffer = tensor.data.f;
+    for (int i = 0; i < keypoints.size(); i++) {
+      for (int j = 0; j < patch_mat[i].rows; ++j) {
+        for (int k = 0; k < patch_mat[i].cols; ++k) {
+          *tensor_buffer++ = patch_mat[i].at<uchar>(j, k) / 128.0f - 1.0f;
+        }
+      }
+    }
+    for (int i = keypoints.size() * kPatchSize * kPatchSize; i < num_bytes / 4;
+         i++) {
+      *tensor_buffer++ = 0;
+    }
+
+    patches->emplace_back(tensor);
+    cc->Outputs().Tag("PATCHES").Add(patches.release(), timestamp);
+  }
+
+  return ::mediapipe::OkStatus();
+}
+
+void FeatureDetectorCalculator::ComputeImagePyramid(
+    const cv::Mat& input_image, std::vector<cv::Mat>* image_pyramid) {
+  cv::Mat tmp_image = input_image;
+  cv::Mat src_image = input_image;
+  for (int i = 0; i < options_.pyramid_level(); ++i) {
+    image_pyramid->push_back(src_image);
+    cv::resize(src_image, tmp_image, cv::Size(), 1.0f / options_.scale_factor(),
+               1.0f / options_.scale_factor());
+    src_image = tmp_image;
+  }
+}
+
+cv::Mat FeatureDetectorCalculator::ExtractPatch(
+    const cv::KeyPoint& feature, const std::vector<cv::Mat>& image_pyramid) {
+  cv::Mat img = image_pyramid[feature.octave];
+  float scale_factor = 1 / pow(options_.scale_factor(), feature.octave);
+  cv::Point2f center =
+      cv::Point2f(feature.pt.x * scale_factor, feature.pt.y * scale_factor);
+  cv::Mat rot = cv::getRotationMatrix2D(center, feature.angle, 1.0);
+  rot.at<double>(0, 2) += kPatchSize / 2 - center.x;
+  rot.at<double>(1, 2) += kPatchSize / 2 - center.y;
+  cv::Mat cropped_img;
+  // perform the affine transformation
+  cv::warpAffine(img, cropped_img, rot, cv::Size(kPatchSize, kPatchSize),
+                 cv::INTER_LINEAR);
+  return cropped_img;
+}
+
+}  // namespace mediapipe
@@ -0,0 +1,24 @@
+// Options for FeatureDetectorCalculator
+syntax = "proto2";
+
+package mediapipe;
+
+import "mediapipe/framework/calculator.proto";
+
+message FeatureDetectorCalculatorOptions {
+  extend CalculatorOptions {
+    optional FeatureDetectorCalculatorOptions ext = 278741680;
+  }
+
+  // Set to true if output patches, otherwise only output cv::KeyPoint
+  optional bool output_patch = 1;
+
+  // The max number of detected features.
+  optional int32 max_features = 2 [default = 200];
+
+  // The number of pyramid levels.
+  optional int32 pyramid_level = 3 [default = 4];
+
+  // Pyramid decimation ratio.
+  optional float scale_factor = 4 [default = 1.2];
+}
@@ -219,8 +219,10 @@ REGISTER_CALCULATOR(ImageCroppingCalculator);
  const auto& input_img = cc->Inputs().Tag(kImageTag).Get<ImageFrame>();
  cv::Mat input_mat = formats::MatView(&input_img);

-  auto [target_width, target_height, rect_center_x, rect_center_y, rotation] =
-      GetCropSpecs(cc, input_img.Width(), input_img.Height());
+  RectSpec specs = GetCropSpecs(cc, input_img.Width(), input_img.Height());
+  int target_width = specs.width, target_height = specs.height,
+      rect_center_x = specs.center_x, rect_center_y = specs.center_y;
+  float rotation = specs.rotation;

  // Get border mode and value for OpenCV.
  int border_mode;
@@ -403,8 +405,10 @@ void ImageCroppingCalculator::GetOutputDimensions(CalculatorContext* cc,
                                                  int src_width, int src_height,
                                                  int* dst_width,
                                                  int* dst_height) {
-  auto [crop_width, crop_height, x_center, y_center, rotation] =
-      GetCropSpecs(cc, src_width, src_height);
+  RectSpec specs = GetCropSpecs(cc, src_width, src_height);
+  int crop_width = specs.width, crop_height = specs.height,
+      x_center = specs.center_x, y_center = specs.center_y;
+  float rotation = specs.rotation;

  const float half_width = crop_width / 2.0f;
  const float half_height = crop_height / 2.0f;
@@ -19,6 +19,11 @@
 #include "mediapipe/gpu/gpu_buffer.h"
 #endif  //  !MEDIAPIPE_DISABLE_GPU

+namespace {
+constexpr char kImageFrameTag[] = "IMAGE";
+constexpr char kGpuBufferTag[] = "IMAGE_GPU";
+}  // namespace
+
 namespace mediapipe {

 // Extracts image properties from the input image and outputs the properties.
@@ -40,13 +45,14 @@ namespace mediapipe {
 class ImagePropertiesCalculator : public CalculatorBase {
 public:
  static ::mediapipe::Status GetContract(CalculatorContract* cc) {
-    RET_CHECK(cc->Inputs().HasTag("IMAGE") ^ cc->Inputs().HasTag("IMAGE_GPU"));
-    if (cc->Inputs().HasTag("IMAGE")) {
-      cc->Inputs().Tag("IMAGE").Set<ImageFrame>();
+    RET_CHECK(cc->Inputs().HasTag(kImageFrameTag) ^
+              cc->Inputs().HasTag(kGpuBufferTag));
+    if (cc->Inputs().HasTag(kImageFrameTag)) {
+      cc->Inputs().Tag(kImageFrameTag).Set<ImageFrame>();
    }
 #if !defined(MEDIAPIPE_DISABLE_GPU)
-    if (cc->Inputs().HasTag("IMAGE_GPU")) {
-      cc->Inputs().Tag("IMAGE_GPU").Set<::mediapipe::GpuBuffer>();
+    if (cc->Inputs().HasTag(kGpuBufferTag)) {
+      cc->Inputs().Tag(kGpuBufferTag).Set<::mediapipe::GpuBuffer>();
    }
 #endif  //  !MEDIAPIPE_DISABLE_GPU

@@ -66,16 +72,17 @@ class ImagePropertiesCalculator : public CalculatorBase {
    int width;
    int height;

-    if (cc->Inputs().HasTag("IMAGE") && !cc->Inputs().Tag("IMAGE").IsEmpty()) {
-      const auto& image = cc->Inputs().Tag("IMAGE").Get<ImageFrame>();
+    if (cc->Inputs().HasTag(kImageFrameTag) &&
+        !cc->Inputs().Tag(kImageFrameTag).IsEmpty()) {
+      const auto& image = cc->Inputs().Tag(kImageFrameTag).Get<ImageFrame>();
      width = image.Width();
      height = image.Height();
    }
 #if !defined(MEDIAPIPE_DISABLE_GPU)
-    if (cc->Inputs().HasTag("IMAGE_GPU") &&
-        !cc->Inputs().Tag("IMAGE_GPU").IsEmpty()) {
+    if (cc->Inputs().HasTag(kGpuBufferTag) &&
+        !cc->Inputs().Tag(kGpuBufferTag).IsEmpty()) {
      const auto& image =
-          cc->Inputs().Tag("IMAGE_GPU").Get<mediapipe::GpuBuffer>();
+          cc->Inputs().Tag(kGpuBufferTag).Get<mediapipe::GpuBuffer>();
      width = image.width();
      height = image.height();
    }
@@ -47,6 +47,9 @@ namespace mediapipe {
 #endif  //  !MEDIAPIPE_DISABLE_GPU

 namespace {
+constexpr char kImageFrameTag[] = "IMAGE";
+constexpr char kGpuBufferTag[] = "IMAGE_GPU";
+
 int RotationModeToDegrees(mediapipe::RotationMode_Mode rotation) {
  switch (rotation) {
    case mediapipe::RotationMode_Mode_UNKNOWN:
@@ -95,7 +98,7 @@ mediapipe::ScaleMode_Mode ParseScaleMode(
 // Scales, rotates, and flips images horizontally or vertically.
 //
 // Input:
-//   One of the following two tags:
+//   One of the following tags:
 //   IMAGE: ImageFrame representing the input image.
 //   IMAGE_GPU: GpuBuffer representing the input image.
 //
@@ -113,7 +116,7 @@ mediapipe::ScaleMode_Mode ParseScaleMode(
 //   corresponding field in the calculator options.
 //
 // Output:
-//   One of the following two tags:
+//   One of the following tags:
 //   IMAGE - ImageFrame representing the output image.
 //   IMAGE_GPU - GpuBuffer representing the output image.
 //
@@ -152,7 +155,8 @@ mediapipe::ScaleMode_Mode ParseScaleMode(
 // Note: To enable horizontal or vertical flipping, specify them in the
 // calculator options. Flipping is applied after rotation.
 //
-// Note: Only scale mode STRETCH is currently supported on CPU.
+// Note: Input defines output, so only matchig types supported:
+// IMAGE -> IMAGE  or  IMAGE_GPU -> IMAGE_GPU
 //
 class ImageTransformationCalculator : public CalculatorBase {
 public:
@@ -186,7 +190,7 @@ class ImageTransformationCalculator : public CalculatorBase {

  bool use_gpu_ = false;
 #if !defined(MEDIAPIPE_DISABLE_GPU)
-  GlCalculatorHelper helper_;
+  GlCalculatorHelper gpu_helper_;
  std::unique_ptr<QuadRenderer> rgb_renderer_;
  std::unique_ptr<QuadRenderer> yuv_renderer_;
  std::unique_ptr<QuadRenderer> ext_rgb_renderer_;
@@ -197,21 +201,22 @@ REGISTER_CALCULATOR(ImageTransformationCalculator);
 // static
 ::mediapipe::Status ImageTransformationCalculator::GetContract(
    CalculatorContract* cc) {
-  RET_CHECK(cc->Inputs().HasTag("IMAGE") ^ cc->Inputs().HasTag("IMAGE_GPU"));
-  RET_CHECK(cc->Outputs().HasTag("IMAGE") ^ cc->Outputs().HasTag("IMAGE_GPU"));
+  // Only one input can be set, and the output type must match.
+  RET_CHECK(cc->Inputs().HasTag(kImageFrameTag) ^
+            cc->Inputs().HasTag(kGpuBufferTag));

  bool use_gpu = false;

-  if (cc->Inputs().HasTag("IMAGE")) {
-    RET_CHECK(cc->Outputs().HasTag("IMAGE"));
-    cc->Inputs().Tag("IMAGE").Set<ImageFrame>();
-    cc->Outputs().Tag("IMAGE").Set<ImageFrame>();
+  if (cc->Inputs().HasTag(kImageFrameTag)) {
+    RET_CHECK(cc->Outputs().HasTag(kImageFrameTag));
+    cc->Inputs().Tag(kImageFrameTag).Set<ImageFrame>();
+    cc->Outputs().Tag(kImageFrameTag).Set<ImageFrame>();
  }
 #if !defined(MEDIAPIPE_DISABLE_GPU)
-  if (cc->Inputs().HasTag("IMAGE_GPU")) {
-    RET_CHECK(cc->Outputs().HasTag("IMAGE_GPU"));
-    cc->Inputs().Tag("IMAGE_GPU").Set<GpuBuffer>();
-    cc->Outputs().Tag("IMAGE_GPU").Set<GpuBuffer>();
+  if (cc->Inputs().HasTag(kGpuBufferTag)) {
+    RET_CHECK(cc->Outputs().HasTag(kGpuBufferTag));
+    cc->Inputs().Tag(kGpuBufferTag).Set<GpuBuffer>();
+    cc->Outputs().Tag(kGpuBufferTag).Set<GpuBuffer>();
    use_gpu |= true;
  }
 #endif  //  !MEDIAPIPE_DISABLE_GPU
@@ -259,7 +264,7 @@ REGISTER_CALCULATOR(ImageTransformationCalculator);

  options_ = cc->Options<ImageTransformationCalculatorOptions>();

-  if (cc->Inputs().HasTag("IMAGE_GPU")) {
+  if (cc->Inputs().HasTag(kGpuBufferTag)) {
    use_gpu_ = true;
  }

@@ -300,7 +305,7 @@ REGISTER_CALCULATOR(ImageTransformationCalculator);
  if (use_gpu_) {
 #if !defined(MEDIAPIPE_DISABLE_GPU)
    // Let the helper access the GL context information.
-    MP_RETURN_IF_ERROR(helper_.Open(cc));
+    MP_RETURN_IF_ERROR(gpu_helper_.Open(cc));
 #else
    RET_CHECK_FAIL() << "GPU processing not enabled.";
 #endif  //  !MEDIAPIPE_DISABLE_GPU
@@ -328,18 +333,14 @@ REGISTER_CALCULATOR(ImageTransformationCalculator);

  if (use_gpu_) {
 #if !defined(MEDIAPIPE_DISABLE_GPU)
-    if (cc->Inputs().Tag("IMAGE_GPU").IsEmpty()) {
-      // Image is missing, hence no way to produce output image. (Timestamp
-      // bound will be updated automatically.)
+    if (cc->Inputs().Tag(kGpuBufferTag).IsEmpty()) {
      return ::mediapipe::OkStatus();
    }
-    return helper_.RunInGlContext(
+    return gpu_helper_.RunInGlContext(
        [this, cc]() -> ::mediapipe::Status { return RenderGpu(cc); });
 #endif  //  !MEDIAPIPE_DISABLE_GPU
  } else {
-    if (cc->Inputs().Tag("IMAGE").IsEmpty()) {
-      // Image is missing, hence no way to produce output image. (Timestamp
-      // bound will be updated automatically.)
+    if (cc->Inputs().Tag(kImageFrameTag).IsEmpty()) {
      return ::mediapipe::OkStatus();
    }
    return RenderCpu(cc);
@@ -354,7 +355,7 @@ REGISTER_CALCULATOR(ImageTransformationCalculator);
    QuadRenderer* rgb_renderer = rgb_renderer_.release();
    QuadRenderer* yuv_renderer = yuv_renderer_.release();
    QuadRenderer* ext_rgb_renderer = ext_rgb_renderer_.release();
-    helper_.RunInGlContext([rgb_renderer, yuv_renderer, ext_rgb_renderer] {
+    gpu_helper_.RunInGlContext([rgb_renderer, yuv_renderer, ext_rgb_renderer] {
      if (rgb_renderer) {
        rgb_renderer->GlTeardown();
        delete rgb_renderer;
@@ -376,18 +377,21 @@ REGISTER_CALCULATOR(ImageTransformationCalculator);

 ::mediapipe::Status ImageTransformationCalculator::RenderCpu(
    CalculatorContext* cc) {
-  int input_width = cc->Inputs().Tag("IMAGE").Get<ImageFrame>().Width();
-  int input_height = cc->Inputs().Tag("IMAGE").Get<ImageFrame>().Height();
+  cv::Mat input_mat;
+  mediapipe::ImageFormat::Format format;

-  const auto& input_img = cc->Inputs().Tag("IMAGE").Get<ImageFrame>();
-  cv::Mat input_mat = formats::MatView(&input_img);
-  cv::Mat scaled_mat;
+  const auto& input = cc->Inputs().Tag(kImageFrameTag).Get<ImageFrame>();
+  input_mat = formats::MatView(&input);
+  format = input.Format();

+  const int input_width = input_mat.cols;
+  const int input_height = input_mat.rows;
  if (!output_height_ || !output_width_) {
    output_height_ = input_height;
    output_width_ = input_width;
  }

+  cv::Mat scaled_mat;
  if (scale_mode_ == mediapipe::ScaleMode_Mode_STRETCH) {
    cv::resize(input_mat, scaled_mat, cv::Size(output_width_, output_height_));
  } else {
@@ -444,10 +448,12 @@ REGISTER_CALCULATOR(ImageTransformationCalculator);
  }

  std::unique_ptr<ImageFrame> output_frame(
-      new ImageFrame(input_img.Format(), output_width, output_height));
+      new ImageFrame(format, output_width, output_height));
  cv::Mat output_mat = formats::MatView(output_frame.get());
  flipped_mat.copyTo(output_mat);
-  cc->Outputs().Tag("IMAGE").Add(output_frame.release(), cc->InputTimestamp());
+  cc->Outputs()
+      .Tag(kImageFrameTag)
+      .Add(output_frame.release(), cc->InputTimestamp());

  return ::mediapipe::OkStatus();
 }
@@ -455,8 +461,9 @@ REGISTER_CALCULATOR(ImageTransformationCalculator);
 ::mediapipe::Status ImageTransformationCalculator::RenderGpu(
    CalculatorContext* cc) {
 #if !defined(MEDIAPIPE_DISABLE_GPU)
-  int input_width = cc->Inputs().Tag("IMAGE_GPU").Get<GpuBuffer>().width();
-  int input_height = cc->Inputs().Tag("IMAGE_GPU").Get<GpuBuffer>().height();
+  const auto& input = cc->Inputs().Tag(kGpuBufferTag).Get<GpuBuffer>();
+  const int input_width = input.width();
+  const int input_height = input.height();

  int output_width;
  int output_height;
@@ -472,7 +479,6 @@ REGISTER_CALCULATOR(ImageTransformationCalculator);
        .Add(padding.release(), cc->InputTimestamp());
  }

-  const auto& input = cc->Inputs().Tag("IMAGE_GPU").Get<GpuBuffer>();
  QuadRenderer* renderer = nullptr;
  GlTexture src1;

@@ -486,11 +492,11 @@ REGISTER_CALCULATOR(ImageTransformationCalculator);
                                 {"video_frame_y", "video_frame_uv"}));
    }
    renderer = yuv_renderer_.get();
-    src1 = helper_.CreateSourceTexture(input, 0);
+    src1 = gpu_helper_.CreateSourceTexture(input, 0);
  } else  // NOLINT(readability/braces)
 #endif    // iOS
  {
-    src1 = helper_.CreateSourceTexture(input);
+    src1 = gpu_helper_.CreateSourceTexture(input);
 #if defined(TEXTURE_EXTERNAL_OES)
    if (src1.target() == GL_TEXTURE_EXTERNAL_OES) {
      if (!ext_rgb_renderer_) {
@@ -516,10 +522,10 @@ REGISTER_CALCULATOR(ImageTransformationCalculator);
  mediapipe::FrameRotation rotation =
      mediapipe::FrameRotationFromDegrees(RotationModeToDegrees(rotation_));

-  auto dst = helper_.CreateDestinationTexture(output_width, output_height,
-                                              input.format());
+  auto dst = gpu_helper_.CreateDestinationTexture(output_width, output_height,
+                                                  input.format());

-  helper_.BindFramebuffer(dst);  // GL_TEXTURE0
+  gpu_helper_.BindFramebuffer(dst);  // GL_TEXTURE0
  glActiveTexture(GL_TEXTURE1);
  glBindTexture(src1.target(), src1.name());

@@ -534,8 +540,8 @@ REGISTER_CALCULATOR(ImageTransformationCalculator);
  // Execute GL commands, before getting result.
  glFlush();

-  auto output = dst.GetFrame<GpuBuffer>();
-  cc->Outputs().Tag("IMAGE_GPU").Add(output.release(), cc->InputTimestamp());
+  auto output = dst.template GetFrame<GpuBuffer>();
+  cc->Outputs().Tag(kGpuBufferTag).Add(output.release(), cc->InputTimestamp());

 #endif  //  !MEDIAPIPE_DISABLE_GPU

@@ -12,6 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.

+#include "mediapipe/calculators/image/opencv_encoded_image_to_image_frame_calculator.pb.h"
 #include "mediapipe/framework/calculator_framework.h"
 #include "mediapipe/framework/formats/image_frame_opencv.h"
 #include "mediapipe/framework/port/opencv_imgcodecs_inc.h"
@@ -34,7 +35,11 @@ namespace mediapipe {
 class OpenCvEncodedImageToImageFrameCalculator : public CalculatorBase {
 public:
  static ::mediapipe::Status GetContract(CalculatorContract* cc);
+  ::mediapipe::Status Open(CalculatorContext* cc) override;
  ::mediapipe::Status Process(CalculatorContext* cc) override;
+
+ private:
+  mediapipe::OpenCvEncodedImageToImageFrameCalculatorOptions options_;
 };

 ::mediapipe::Status OpenCvEncodedImageToImageFrameCalculator::GetContract(
@@ -44,13 +49,29 @@ class OpenCvEncodedImageToImageFrameCalculator : public CalculatorBase {
  return ::mediapipe::OkStatus();
 }

+::mediapipe::Status OpenCvEncodedImageToImageFrameCalculator::Open(
+    CalculatorContext* cc) {
+  options_ =
+      cc->Options<mediapipe::OpenCvEncodedImageToImageFrameCalculatorOptions>();
+  return ::mediapipe::OkStatus();
+}
+
 ::mediapipe::Status OpenCvEncodedImageToImageFrameCalculator::Process(
    CalculatorContext* cc) {
  const std::string& contents = cc->Inputs().Index(0).Get<std::string>();
  const std::vector<char> contents_vector(contents.begin(), contents.end());
-  cv::Mat decoded_mat =
-      cv::imdecode(contents_vector, -1 /* return the loaded image as-is */);
-
+  cv::Mat decoded_mat;
+  if (options_.apply_orientation_from_exif_data()) {
+    // We want to respect the orientation from the EXIF data, which
+    // IMREAD_UNCHANGED ignores, but otherwise we want to be as permissive as
+    // possible with our reading flags. Therefore, we use IMREAD_ANYCOLOR and
+    // IMREAD_ANYDEPTH.
+    decoded_mat = cv::imdecode(contents_vector,
+                               cv::IMREAD_ANYCOLOR | cv::IMREAD_ANYDEPTH);
+  } else {
+    // Return the loaded image as-is
+    decoded_mat = cv::imdecode(contents_vector, cv::IMREAD_UNCHANGED);
+  }
  ImageFormat::Format image_format = ImageFormat::UNKNOWN;
  cv::Mat output_mat;
  switch (decoded_mat.channels()) {
@@ -70,7 +91,8 @@ class OpenCvEncodedImageToImageFrameCalculator : public CalculatorBase {
             << "Unsupported number of channels: " << decoded_mat.channels();
  }
  std::unique_ptr<ImageFrame> output_frame = absl::make_unique<ImageFrame>(
-      image_format, decoded_mat.size().width, decoded_mat.size().height);
+      image_format, decoded_mat.size().width, decoded_mat.size().height,
+      ImageFrame::kGlDefaultAlignmentBoundary);
  output_mat.copyTo(formats::MatView(output_frame.get()));
  cc->Outputs().Index(0).Add(output_frame.release(), cc->InputTimestamp());
  return ::mediapipe::OkStatus();
@@ -0,0 +1,30 @@
+// Copyright 2020 The MediaPipe Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+syntax = "proto2";
+
+package mediapipe;
+
+import "mediapipe/framework/calculator.proto";
+
+message OpenCvEncodedImageToImageFrameCalculatorOptions {
+  extend CalculatorOptions {
+    optional OpenCvEncodedImageToImageFrameCalculatorOptions ext = 303447308;
+  }
+
+  // If set, we will attempt to automatically apply the orientation specified by
+  // the image's EXIF data when loading the image. Otherwise, the image data
+  // will be loaded as-is.
+  optional bool apply_orientation_from_exif_data = 1 [default = false];
+}
@@ -17,6 +17,9 @@
 #include "mediapipe/calculators/image/recolor_calculator.pb.h"
 #include "mediapipe/framework/calculator_framework.h"
 #include "mediapipe/framework/formats/image_frame.h"
+#include "mediapipe/framework/formats/image_frame_opencv.h"
+#include "mediapipe/framework/port/opencv_core_inc.h"
+#include "mediapipe/framework/port/opencv_imgproc_inc.h"
 #include "mediapipe/framework/port/ret_check.h"
 #include "mediapipe/framework/port/status.h"
 #include "mediapipe/util/color.pb.h"
@@ -29,6 +32,11 @@

 namespace {
 enum { ATTRIB_VERTEX, ATTRIB_TEXTURE_POSITION, NUM_ATTRIBUTES };
+
+constexpr char kImageFrameTag[] = "IMAGE";
+constexpr char kMaskCpuTag[] = "MASK";
+constexpr char kGpuBufferTag[] = "IMAGE_GPU";
+constexpr char kMaskGpuTag[] = "MASK_GPU";
 }  // namespace

 namespace mediapipe {
@@ -39,8 +47,6 @@ namespace mediapipe {
 // The luminance of the input image is used to adjust the blending weight,
 // to help preserve image textures.
 //
-// TODO implement cpu support.
-//
 // Inputs:
 //   One of the following IMAGE tags:
 //   IMAGE: An ImageFrame input image, RGB or RGBA.
@@ -71,6 +77,8 @@ namespace mediapipe {
 //    }
 //  }
 //
+// Note: Cannot mix-match CPU & GPU inputs/outputs.
+//       CPU-in & CPU-out <or> GPU-in & GPU-out
 class RecolorCalculator : public CalculatorBase {
 public:
  RecolorCalculator() = default;
@@ -109,35 +117,42 @@ REGISTER_CALCULATOR(RecolorCalculator);
  bool use_gpu = false;

 #if !defined(MEDIAPIPE_DISABLE_GPU)
-  if (cc->Inputs().HasTag("IMAGE_GPU")) {
-    cc->Inputs().Tag("IMAGE_GPU").Set<mediapipe::GpuBuffer>();
+  if (cc->Inputs().HasTag(kGpuBufferTag)) {
+    cc->Inputs().Tag(kGpuBufferTag).Set<mediapipe::GpuBuffer>();
    use_gpu |= true;
  }
 #endif  //  !MEDIAPIPE_DISABLE_GPU
-  if (cc->Inputs().HasTag("IMAGE")) {
-    cc->Inputs().Tag("IMAGE").Set<ImageFrame>();
+  if (cc->Inputs().HasTag(kImageFrameTag)) {
+    cc->Inputs().Tag(kImageFrameTag).Set<ImageFrame>();
  }

 #if !defined(MEDIAPIPE_DISABLE_GPU)
-  if (cc->Inputs().HasTag("MASK_GPU")) {
-    cc->Inputs().Tag("MASK_GPU").Set<mediapipe::GpuBuffer>();
+  if (cc->Inputs().HasTag(kMaskGpuTag)) {
+    cc->Inputs().Tag(kMaskGpuTag).Set<mediapipe::GpuBuffer>();
    use_gpu |= true;
  }
 #endif  //  !MEDIAPIPE_DISABLE_GPU
-  if (cc->Inputs().HasTag("MASK")) {
-    cc->Inputs().Tag("MASK").Set<ImageFrame>();
+  if (cc->Inputs().HasTag(kMaskCpuTag)) {
+    cc->Inputs().Tag(kMaskCpuTag).Set<ImageFrame>();
  }

 #if !defined(MEDIAPIPE_DISABLE_GPU)
-  if (cc->Outputs().HasTag("IMAGE_GPU")) {
-    cc->Outputs().Tag("IMAGE_GPU").Set<mediapipe::GpuBuffer>();
+  if (cc->Outputs().HasTag(kGpuBufferTag)) {
+    cc->Outputs().Tag(kGpuBufferTag).Set<mediapipe::GpuBuffer>();
    use_gpu |= true;
  }
 #endif  //  !MEDIAPIPE_DISABLE_GPU
-  if (cc->Outputs().HasTag("IMAGE")) {
-    cc->Outputs().Tag("IMAGE").Set<ImageFrame>();
+  if (cc->Outputs().HasTag(kImageFrameTag)) {
+    cc->Outputs().Tag(kImageFrameTag).Set<ImageFrame>();
  }

+  // Confirm only one of the input streams is present.
+  RET_CHECK(cc->Inputs().HasTag(kImageFrameTag) ^
+            cc->Inputs().HasTag(kGpuBufferTag));
+  // Confirm only one of the output streams is present.
+  RET_CHECK(cc->Outputs().HasTag(kImageFrameTag) ^
+            cc->Outputs().HasTag(kGpuBufferTag));
+
  if (use_gpu) {
 #if !defined(MEDIAPIPE_DISABLE_GPU)
    MP_RETURN_IF_ERROR(mediapipe::GlCalculatorHelper::UpdateContract(cc));
@@ -150,7 +165,7 @@ REGISTER_CALCULATOR(RecolorCalculator);
 ::mediapipe::Status RecolorCalculator::Open(CalculatorContext* cc) {
  cc->SetOffset(TimestampDiff(0));

-  if (cc->Inputs().HasTag("IMAGE_GPU")) {
+  if (cc->Inputs().HasTag(kGpuBufferTag)) {
    use_gpu_ = true;
 #if !defined(MEDIAPIPE_DISABLE_GPU)
    MP_RETURN_IF_ERROR(gpu_helper_.Open(cc));
@@ -193,17 +208,74 @@ REGISTER_CALCULATOR(RecolorCalculator);
 }

 ::mediapipe::Status RecolorCalculator::RenderCpu(CalculatorContext* cc) {
-  return ::mediapipe::UnimplementedError("CPU support is not implemented yet.");
+  if (cc->Inputs().Tag(kMaskCpuTag).IsEmpty()) {
+    return ::mediapipe::OkStatus();
+  }
+  // Get inputs and setup output.
+  const auto& input_img = cc->Inputs().Tag(kImageFrameTag).Get<ImageFrame>();
+  const auto& mask_img = cc->Inputs().Tag(kMaskCpuTag).Get<ImageFrame>();
+
+  cv::Mat input_mat = formats::MatView(&input_img);
+  cv::Mat mask_mat = formats::MatView(&mask_img);
+
+  RET_CHECK(input_mat.channels() == 3);  // RGB only.
+
+  if (mask_mat.channels() > 1) {
+    std::vector<cv::Mat> channels;
+    cv::split(mask_mat, channels);
+    if (mask_channel_ == mediapipe::RecolorCalculatorOptions_MaskChannel_ALPHA)
+      mask_mat = channels[3];
+    else
+      mask_mat = channels[0];
+  }
+  cv::Mat mask_full;
+  cv::resize(mask_mat, mask_full, input_mat.size());
+
+  auto output_img = absl::make_unique<ImageFrame>(
+      input_img.Format(), input_mat.cols, input_mat.rows);
+  cv::Mat output_mat = mediapipe::formats::MatView(output_img.get());
+
+  // From GPU shader:
+  /*
+      vec4 weight = texture2D(mask, sample_coordinate);
+      vec4 color1 = texture2D(frame, sample_coordinate);
+      vec4 color2 = vec4(recolor, 1.0);
+
+      float luminance = dot(color1.rgb, vec3(0.299, 0.587, 0.114));
+      float mix_value = weight.MASK_COMPONENT * luminance;
+
+      fragColor = mix(color1, color2, mix_value);
+  */
+  for (int i = 0; i < output_mat.rows; ++i) {
+    for (int j = 0; j < output_mat.cols; ++j) {
+      float weight = mask_full.at<uchar>(i, j) * (1.0 / 255.0);
+      cv::Vec3f color1 = input_mat.at<cv::Vec3b>(i, j);
+      cv::Vec3f color2 = {color_[0], color_[1], color_[2]};
+
+      float luminance =
+          (color1[0] * 0.299 + color1[1] * 0.587 + color1[2] * 0.114) / 255;
+      float mix_value = weight * luminance;
+
+      cv::Vec3b mix_color = color1 * (1.0 - mix_value) + color2 * mix_value;
+      output_mat.at<cv::Vec3b>(i, j) = mix_color;
+    }
+  }
+
+  cc->Outputs()
+      .Tag(kImageFrameTag)
+      .Add(output_img.release(), cc->InputTimestamp());
+
+  return ::mediapipe::OkStatus();
 }

 ::mediapipe::Status RecolorCalculator::RenderGpu(CalculatorContext* cc) {
-  if (cc->Inputs().Tag("MASK_GPU").IsEmpty()) {
+  if (cc->Inputs().Tag(kMaskGpuTag).IsEmpty()) {
    return ::mediapipe::OkStatus();
  }
 #if !defined(MEDIAPIPE_DISABLE_GPU)
  // Get inputs and setup output.
-  const Packet& input_packet = cc->Inputs().Tag("IMAGE_GPU").Value();
-  const Packet& mask_packet = cc->Inputs().Tag("MASK_GPU").Value();
+  const Packet& input_packet = cc->Inputs().Tag(kGpuBufferTag).Value();
+  const Packet& mask_packet = cc->Inputs().Tag(kMaskGpuTag).Value();

  const auto& input_buffer = input_packet.Get<mediapipe::GpuBuffer>();
  const auto& mask_buffer = mask_packet.Get<mediapipe::GpuBuffer>();
@@ -233,7 +305,7 @@ REGISTER_CALCULATOR(RecolorCalculator);

  // Send result image in GPU packet.
  auto output = dst_tex.GetFrame<mediapipe::GpuBuffer>();
-  cc->Outputs().Tag("IMAGE_GPU").Add(output.release(), cc->InputTimestamp());
+  cc->Outputs().Tag(kGpuBufferTag).Add(output.release(), cc->InputTimestamp());

  // Cleanup
  img_tex.Release();
@@ -303,9 +375,9 @@ void RecolorCalculator::GlRender() {

  if (!options.has_color()) RET_CHECK_FAIL() << "Missing color option.";

-  color_.push_back(options.color().r() / 255.0);
-  color_.push_back(options.color().g() / 255.0);
-  color_.push_back(options.color().b() / 255.0);
+  color_.push_back(options.color().r());
+  color_.push_back(options.color().g());
+  color_.push_back(options.color().b());

  return ::mediapipe::OkStatus();
 }
@@ -378,8 +450,8 @@ void RecolorCalculator::GlRender() {
  glUseProgram(program_);
  glUniform1i(glGetUniformLocation(program_, "frame"), 1);
  glUniform1i(glGetUniformLocation(program_, "mask"), 2);
-  glUniform3f(glGetUniformLocation(program_, "recolor"), color_[0], color_[1],
-              color_[2]);
+  glUniform3f(glGetUniformLocation(program_, "recolor"), color_[0] / 255.0,
+              color_[1] / 255.0, color_[2] / 255.0);
 #endif  //  !MEDIAPIPE_DISABLE_GPU

  return ::mediapipe::OkStatus();
@@ -1110,6 +1110,7 @@ cc_test(
        ],
        "//mediapipe:android": [
            "@org_tensorflow//tensorflow/core:android_tensorflow_lib_with_ops_lite_proto_no_rtti_lib",
+            "@org_tensorflow//tensorflow/core:android_tensorflow_test_lib",
        ],
        "//mediapipe:ios": [
            "@org_tensorflow//tensorflow/core:ios_tensorflow_test_lib",
@@ -244,7 +244,7 @@ class PackMediaSequenceCalculator : public CalculatorBase {
  ::mediapipe::Status VerifySequence() {
    std::string error_msg = "Missing features - ";
    bool all_present = true;
-    for (auto iter : features_present_) {
+    for (const auto& iter : features_present_) {
      if (!iter.second) {
        all_present = false;
        absl::StrAppend(&error_msg, iter.first, ", ");
@@ -51,10 +51,11 @@ static constexpr char kStringSavedModelPath[] = "STRING_SAVED_MODEL_PATH";
 #endif
 }

-// If options.convert_signature_to_tags() will convert letters to uppercase
-// and replace /'s with _'s. If set, this enables the standard SavedModel
-// classification, regression, and prediction signatures to be used as
-// uppercase INPUTS and OUTPUTS tags for streams.
+// If options.convert_signature_to_tags() is set, will convert letters to
+// uppercase and replace /'s and -'s with _'s. This enables the standard
+// SavedModel classification, regression, and prediction signatures to be used
+// as uppercase INPUTS and OUTPUTS tags for streams and supports other common
+// patterns.
 const std::string MaybeConvertSignatureToTag(
    const std::string& name,
    const TensorFlowSessionFromSavedModelCalculatorOptions& options) {
@@ -64,6 +65,7 @@ const std::string MaybeConvertSignatureToTag(
    std::transform(name.begin(), name.end(), output.begin(),
                   [](unsigned char c) { return std::toupper(c); });
    output = absl::StrReplaceAll(output, {{"/", "_"}});
+    output = absl::StrReplaceAll(output, {{"-", "_"}});
    return output;
  } else {
    return name;
@@ -124,7 +126,7 @@ class TensorFlowSessionFromSavedModelCalculator : public CalculatorBase {
    // Set user specified tags properly.
    // If no tags specified will use tensorflow::kSavedModelTagServe by default.
    std::unordered_set<std::string> tags_set;
-    for (std::string tag : options.saved_model_tag()) {
+    for (const std::string& tag : options.saved_model_tag()) {
      tags_set.insert(tag);
    }
    if (tags_set.empty()) {
@@ -32,8 +32,8 @@ message TensorFlowSessionFromSavedModelCalculatorOptions {
  // The name of the generic signature to load into the mapping from tags to
  // tensor names.
  optional string signature_name = 2 [default = "serving_default"];
-  // Whether to convert the signature keys to uppercase and switch /'s to
-  // _'s, which enables standard signatures to be used as Tags.
+  // Whether to convert the signature keys to uppercase as well as switch /'s
+  // and -'s to _'s, which enables common signatures to be used as Tags.
  optional bool convert_signature_to_tags = 3 [default = true];
  // If true, saved_model_path can have multiple exported models in
  // subdirectories saved_model_path/%08d and the alphabetically last (i.e.,
@@ -53,10 +53,11 @@ static constexpr char kStringSavedModelPath[] = "STRING_SAVED_MODEL_PATH";
 #endif
 }

-// If options.convert_signature_to_tags() will convert letters to uppercase
-// and replace /'s with _'s. If set, this enables the standard SavedModel
-// classification, regression, and prediction signatures to be used as
-// uppercase INPUTS and OUTPUTS tags for streams.
+// If options.convert_signature_to_tags() is set, will convert letters to
+// uppercase and replace /'s and -'s with _'s. This enables the standard
+// SavedModel classification, regression, and prediction signatures to be used
+// as uppercase INPUTS and OUTPUTS tags for streams and supports other common
+// patterns.
 const std::string MaybeConvertSignatureToTag(
    const std::string& name,
    const TensorFlowSessionFromSavedModelGeneratorOptions& options) {
@@ -66,6 +67,7 @@ const std::string MaybeConvertSignatureToTag(
    std::transform(name.begin(), name.end(), output.begin(),
                   [](unsigned char c) { return std::toupper(c); });
    output = absl::StrReplaceAll(output, {{"/", "_"}});
+    output = absl::StrReplaceAll(output, {{"-", "_"}});
    return output;
  } else {
    return name;
@@ -119,7 +121,7 @@ class TensorFlowSessionFromSavedModelGenerator : public PacketGenerator {
    // Set user specified tags properly.
    // If no tags specified will use tensorflow::kSavedModelTagServe by default.
    std::unordered_set<std::string> tags_set;
-    for (std::string tag : options.saved_model_tag()) {
+    for (const std::string& tag : options.saved_model_tag()) {
      tags_set.insert(tag);
    }
    if (tags_set.empty()) {
@@ -32,8 +32,8 @@ message TensorFlowSessionFromSavedModelGeneratorOptions {
  // The name of the generic signature to load into the mapping from tags to
  // tensor names.
  optional string signature_name = 2 [default = "serving_default"];
-  // Whether to convert the signature keys to uppercase and switch /'s to
-  // _'s, which enables standard signatures to be used as Tags.
+  // Whether to convert the signature keys to uppercase as well as switch /'s
+  // and -'s to _'s, which enables common signatures to be used as Tags.
  optional bool convert_signature_to_tags = 3 [default = true];
  // If true, saved_model_path can have multiple exported models in
  // subdirectories saved_model_path/%08d and the alphabetically last (i.e.,
@@ -222,9 +222,11 @@ cc_library(
    deps = [
        ":util",
        ":tflite_inference_calculator_cc_proto",
+        "@com_google_absl//absl/memory",
        "//mediapipe/framework:calculator_framework",
        "//mediapipe/util:resource_util",
        "@org_tensorflow//tensorflow/lite:framework",
+        "@org_tensorflow//tensorflow/lite/delegates/xnnpack:xnnpack_delegate",
        "@org_tensorflow//tensorflow/lite/kernels:builtin_ops",
        "//mediapipe/framework/stream_handler:fixed_size_input_stream_handler",
        "//mediapipe/framework/port:ret_check",
@@ -241,6 +243,7 @@ cc_library(
            "@org_tensorflow//tensorflow/lite/delegates/gpu:metal_delegate_internal",
        ],
        "//conditions:default": [
+            "//mediapipe/util/tflite:tflite_gpu_runner",
            "//mediapipe/gpu:gl_calculator_helper",
            "//mediapipe/gpu:gpu_buffer",
            "@org_tensorflow//tensorflow/lite/delegates/gpu/common:shape",
@@ -254,6 +257,10 @@ cc_library(
        "//mediapipe:android": [
            "@org_tensorflow//tensorflow/lite/delegates/nnapi:nnapi_delegate",
        ],
+    }) + select({
+        "//conditions:default": [
+            "//mediapipe/util:cpu_util",
+        ],
    }),
    alwayslink = 1,
 )
@@ -308,6 +315,20 @@ cc_library(
    alwayslink = 1,
 )

+cc_library(
+    name = "tflite_model_calculator",
+    srcs = ["tflite_model_calculator.cc"],
+    visibility = ["//visibility:public"],
+    deps = [
+        ":util",
+        "//mediapipe/framework:calculator_framework",
+        "//mediapipe/framework:packet",
+        "//mediapipe/framework/port:ret_check",
+        "@org_tensorflow//tensorflow/lite:framework",
+    ],
+    alwayslink = 1,
+)
+
 cc_library(
    name = "tflite_tensors_to_segmentation_calculator",
    srcs = ["tflite_tensors_to_segmentation_calculator.cc"],
@@ -431,7 +452,7 @@ cc_library(
        "//mediapipe:android": [
            "//mediapipe/util/android/file/base",
        ],
-        "//mediapipe:apple": [
+        "//mediapipe:ios": [
            "//mediapipe/util/android/file/base",
        ],
        "//mediapipe:macos": [
@@ -478,6 +499,9 @@ cc_test(
    deps = [
        ":tflite_inference_calculator",
        ":tflite_inference_calculator_cc_proto",
+        ":tflite_model_calculator",
+        "//mediapipe/calculators/core:constant_side_packet_calculator",
+        "//mediapipe/calculators/util:local_file_contents_calculator",
        "//mediapipe/framework:calculator_framework",
        "//mediapipe/framework:calculator_runner",
        "//mediapipe/framework/deps:file_path",
@@ -485,7 +509,9 @@ cc_test(
        "//mediapipe/framework/port:integral_types",
        "//mediapipe/framework/port:parse_text_proto",
        "//mediapipe/framework/tool:validate_type",
+        "@com_google_absl//absl/memory",
        "@com_google_absl//absl/strings",
+        "@com_google_absl//absl/types:optional",
        "@org_tensorflow//tensorflow/lite:framework",
        "@org_tensorflow//tensorflow/lite/kernels:builtin_ops",
    ],
@@ -511,3 +537,19 @@ cc_test(
        "@org_tensorflow//tensorflow/lite/kernels:builtin_ops",
    ],
 )
+
+cc_test(
+    name = "tflite_model_calculator_test",
+    srcs = ["tflite_model_calculator_test.cc"],
+    data = ["testdata/add.bin"],
+    deps = [
+        ":tflite_model_calculator",
+        "//mediapipe/calculators/core:constant_side_packet_calculator",
+        "//mediapipe/calculators/util:local_file_contents_calculator",
+        "//mediapipe/framework:calculator_framework",
+        "//mediapipe/framework:calculator_runner",
+        "//mediapipe/framework/port:gtest_main",
+        "//mediapipe/framework/port:parse_text_proto",
+        "@org_tensorflow//tensorflow/lite:framework",
+    ],
+)
@@ -63,6 +63,10 @@ typedef Eigen::Matrix<float, Eigen::Dynamic, Eigen::Dynamic, Eigen::RowMajor>
 typedef Eigen::Matrix<float, Eigen::Dynamic, Eigen::Dynamic, Eigen::ColMajor>
    ColMajorMatrixXf;

+constexpr char kImageFrameTag[] = "IMAGE";
+constexpr char kGpuBufferTag[] = "IMAGE_GPU";
+constexpr char kTensorsTag[] = "TENSORS";
+constexpr char kTensorsGpuTag[] = "TENSORS_GPU";
 }  // namespace

 namespace mediapipe {
@@ -124,6 +128,9 @@ struct GPUData {
 //  GPU tensors are currently only supported on mobile platforms.
 //  This calculator uses FixedSizeInputStreamHandler by default.
 //
+// Note: Input defines output, so only these type sets are supported:
+// IMAGE -> TENSORS | IMAGE_GPU -> TENSORS_GPU | MATRIX -> TENSORS
+//
 class TfLiteConverterCalculator : public CalculatorBase {
 public:
  static ::mediapipe::Status GetContract(CalculatorContract* cc);
@@ -138,9 +145,9 @@ class TfLiteConverterCalculator : public CalculatorBase {
  template <class T>
  ::mediapipe::Status NormalizeImage(const ImageFrame& image_frame,
                                     bool zero_center, bool flip_vertically,
-                                     float* tensor_buffer);
+                                     float* tensor_ptr);
  ::mediapipe::Status CopyMatrixToTensor(const Matrix& matrix,
-                                         float* tensor_buffer);
+                                         float* tensor_ptr);
  ::mediapipe::Status ProcessCPU(CalculatorContext* cc);
  ::mediapipe::Status ProcessGPU(CalculatorContext* cc);

@@ -166,33 +173,35 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);

 ::mediapipe::Status TfLiteConverterCalculator::GetContract(
    CalculatorContract* cc) {
-  const bool has_image_tag = cc->Inputs().HasTag("IMAGE");
-  const bool has_image_gpu_tag = cc->Inputs().HasTag("IMAGE_GPU");
-  const bool has_matrix_tag = cc->Inputs().HasTag("MATRIX");
  // Confirm only one of the input streams is present.
-  RET_CHECK(has_image_tag ^ has_image_gpu_tag ^ has_matrix_tag &&
-            !(has_image_tag && has_image_gpu_tag && has_matrix_tag));
+  RET_CHECK(cc->Inputs().HasTag(kImageFrameTag) ^
+            cc->Inputs().HasTag(kGpuBufferTag) ^ cc->Inputs().HasTag("MATRIX"));

  // Confirm only one of the output streams is present.
-  RET_CHECK(cc->Outputs().HasTag("TENSORS") ^
-            cc->Outputs().HasTag("TENSORS_GPU"));
+  RET_CHECK(cc->Outputs().HasTag(kTensorsTag) ^
+            cc->Outputs().HasTag(kTensorsGpuTag));

  bool use_gpu = false;

-  if (cc->Inputs().HasTag("IMAGE")) cc->Inputs().Tag("IMAGE").Set<ImageFrame>();
-  if (cc->Inputs().HasTag("MATRIX")) cc->Inputs().Tag("MATRIX").Set<Matrix>();
+  if (cc->Inputs().HasTag(kImageFrameTag)) {
+    cc->Inputs().Tag(kImageFrameTag).Set<ImageFrame>();
+  }
+  if (cc->Inputs().HasTag("MATRIX")) {
+    cc->Inputs().Tag("MATRIX").Set<Matrix>();
+  }
 #if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__)
-  if (cc->Inputs().HasTag("IMAGE_GPU")) {
-    cc->Inputs().Tag("IMAGE_GPU").Set<mediapipe::GpuBuffer>();
+  if (cc->Inputs().HasTag(kGpuBufferTag)) {
+    cc->Inputs().Tag(kGpuBufferTag).Set<mediapipe::GpuBuffer>();
    use_gpu |= true;
  }
 #endif  //  !MEDIAPIPE_DISABLE_GPU

-  if (cc->Outputs().HasTag("TENSORS"))
-    cc->Outputs().Tag("TENSORS").Set<std::vector<TfLiteTensor>>();
+  if (cc->Outputs().HasTag(kTensorsTag)) {
+    cc->Outputs().Tag(kTensorsTag).Set<std::vector<TfLiteTensor>>();
+  }
 #if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__)
-  if (cc->Outputs().HasTag("TENSORS_GPU")) {
-    cc->Outputs().Tag("TENSORS_GPU").Set<std::vector<GpuTensor>>();
+  if (cc->Outputs().HasTag(kTensorsGpuTag)) {
+    cc->Outputs().Tag(kTensorsGpuTag).Set<std::vector<GpuTensor>>();
    use_gpu |= true;
  }
 #endif  //  !MEDIAPIPE_DISABLE_GPU
@@ -216,8 +225,8 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);

  MP_RETURN_IF_ERROR(LoadOptions(cc));

-  if (cc->Inputs().HasTag("IMAGE_GPU") ||
-      cc->Outputs().HasTag("IMAGE_OUT_GPU")) {
+  if (cc->Inputs().HasTag(kGpuBufferTag) ||
+      cc->Outputs().HasTag(kGpuBufferTag)) {
 #if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__)
    use_gpu_ = true;
 #else
@@ -227,8 +236,8 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);

  if (use_gpu_) {
    // Cannot mix CPU/GPU streams.
-    RET_CHECK(cc->Inputs().HasTag("IMAGE_GPU") &&
-              cc->Outputs().HasTag("TENSORS_GPU"));
+    RET_CHECK(cc->Inputs().HasTag(kGpuBufferTag) &&
+              cc->Outputs().HasTag(kTensorsGpuTag));
    // Cannot use quantization.
    use_quantized_tensors_ = false;
 #if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
@@ -248,7 +257,6 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);

 ::mediapipe::Status TfLiteConverterCalculator::Process(CalculatorContext* cc) {
  if (use_gpu_) {
-    // GpuBuffer to tflite::gpu::GlBuffer conversion.
    if (!initialized_) {
      MP_RETURN_IF_ERROR(InitGpu(cc));
      initialized_ = true;
@@ -259,7 +267,6 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);
    // Convert to CPU tensors or Matrix type.
    MP_RETURN_IF_ERROR(ProcessCPU(cc));
  }
-
  return ::mediapipe::OkStatus();
 }

@@ -275,24 +282,26 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);

 ::mediapipe::Status TfLiteConverterCalculator::ProcessCPU(
    CalculatorContext* cc) {
-  if (cc->Inputs().HasTag("IMAGE")) {
+  if (cc->Inputs().HasTag(kImageFrameTag)) {
    // CPU ImageFrame to TfLiteTensor conversion.

-    const auto& image_frame = cc->Inputs().Tag("IMAGE").Get<ImageFrame>();
+    const auto& image_frame =
+        cc->Inputs().Tag(kImageFrameTag).Get<ImageFrame>();
    const int height = image_frame.Height();
    const int width = image_frame.Width();
    const int channels = image_frame.NumberOfChannels();
    const int channels_preserved = std::min(channels, max_num_channels_);
+    const mediapipe::ImageFormat::Format format = image_frame.Format();

    if (!initialized_) {
-      if (!(image_frame.Format() == mediapipe::ImageFormat::SRGBA ||
-            image_frame.Format() == mediapipe::ImageFormat::SRGB ||
-            image_frame.Format() == mediapipe::ImageFormat::GRAY8 ||
-            image_frame.Format() == mediapipe::ImageFormat::VEC32F1))
+      if (!(format == mediapipe::ImageFormat::SRGBA ||
+            format == mediapipe::ImageFormat::SRGB ||
+            format == mediapipe::ImageFormat::GRAY8 ||
+            format == mediapipe::ImageFormat::VEC32F1))
        RET_CHECK_FAIL() << "Unsupported CPU input format.";
      TfLiteQuantization quant;
      if (use_quantized_tensors_) {
-        RET_CHECK(image_frame.Format() != mediapipe::ImageFormat::VEC32F1)
+        RET_CHECK(format != mediapipe::ImageFormat::VEC32F1)
            << "Only 8-bit input images are supported for quantization.";
        quant.type = kTfLiteAffineQuantization;
        quant.params = nullptr;
@@ -349,8 +358,9 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);

    auto output_tensors = absl::make_unique<std::vector<TfLiteTensor>>();
    output_tensors->emplace_back(*tensor);
-    cc->Outputs().Tag("TENSORS").Add(output_tensors.release(),
-                                     cc->InputTimestamp());
+    cc->Outputs()
+        .Tag(kTensorsTag)
+        .Add(output_tensors.release(), cc->InputTimestamp());
  } else if (cc->Inputs().HasTag("MATRIX")) {
    // CPU Matrix to TfLiteTensor conversion.

@@ -371,15 +381,16 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);
    interpreter_->ResizeInputTensor(tensor_idx, {height, width, channels});
    interpreter_->AllocateTensors();

-    float* tensor_buffer = tensor->data.f;
-    RET_CHECK(tensor_buffer);
+    float* tensor_ptr = tensor->data.f;
+    RET_CHECK(tensor_ptr);

-    MP_RETURN_IF_ERROR(CopyMatrixToTensor(matrix, tensor_buffer));
+    MP_RETURN_IF_ERROR(CopyMatrixToTensor(matrix, tensor_ptr));

    auto output_tensors = absl::make_unique<std::vector<TfLiteTensor>>();
    output_tensors->emplace_back(*tensor);
-    cc->Outputs().Tag("TENSORS").Add(output_tensors.release(),
-                                     cc->InputTimestamp());
+    cc->Outputs()
+        .Tag(kTensorsTag)
+        .Add(output_tensors.release(), cc->InputTimestamp());
  }

  return ::mediapipe::OkStatus();
@@ -389,7 +400,8 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);
    CalculatorContext* cc) {
 #if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
  // GpuBuffer to tflite::gpu::GlBuffer conversion.
-  const auto& input = cc->Inputs().Tag("IMAGE_GPU").Get<mediapipe::GpuBuffer>();
+  const auto& input =
+      cc->Inputs().Tag(kGpuBufferTag).Get<mediapipe::GpuBuffer>();
  MP_RETURN_IF_ERROR(
      gpu_helper_.RunInGlContext([this, &input]() -> ::mediapipe::Status {
        // Convert GL texture into TfLite GlBuffer (SSBO).
@@ -421,11 +433,12 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);
        return ::mediapipe::OkStatus();
      }));
  cc->Outputs()
-      .Tag("TENSORS_GPU")
+      .Tag(kTensorsGpuTag)
      .Add(output_tensors.release(), cc->InputTimestamp());
 #elif defined(MEDIAPIPE_IOS)
  // GpuBuffer to id<MTLBuffer> conversion.
-  const auto& input = cc->Inputs().Tag("IMAGE_GPU").Get<mediapipe::GpuBuffer>();
+  const auto& input =
+      cc->Inputs().Tag(kGpuBufferTag).Get<mediapipe::GpuBuffer>();
  id<MTLCommandBuffer> command_buffer = [gpu_helper_ commandBuffer];

  id<MTLTexture> src_texture = [gpu_helper_ metalTextureWithGpuBuffer:input];
@@ -457,7 +470,7 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);
                    commandBuffer:command_buffer];

  cc->Outputs()
-      .Tag("TENSORS_GPU")
+      .Tag(kTensorsGpuTag)
      .Add(output_tensors.release(), cc->InputTimestamp());
 #else
  RET_CHECK_FAIL() << "GPU processing is not enabled.";
@@ -469,7 +482,8 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);
 ::mediapipe::Status TfLiteConverterCalculator::InitGpu(CalculatorContext* cc) {
 #if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__)
  // Get input image sizes.
-  const auto& input = cc->Inputs().Tag("IMAGE_GPU").Get<mediapipe::GpuBuffer>();
+  const auto& input =
+      cc->Inputs().Tag(kGpuBufferTag).Get<mediapipe::GpuBuffer>();
  mediapipe::ImageFormat::Format format =
      mediapipe::ImageFormatForGpuBufferFormat(input.format());
  gpu_data_out_ = absl::make_unique<GPUData>();
@@ -612,7 +626,7 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);
  CHECK_LE(max_num_channels_, 4);
  CHECK_NE(max_num_channels_, 2);
 #if defined(MEDIAPIPE_IOS)
-  if (cc->Inputs().HasTag("IMAGE_GPU"))
+  if (cc->Inputs().HasTag(kGpuBufferTag))
    // Currently on iOS, tflite gpu input tensor must be 4 channels,
    // so input image must be 4 channels also (checked in InitGpu).
    max_num_channels_ = 4;
@@ -627,7 +641,7 @@ REGISTER_CALCULATOR(TfLiteConverterCalculator);
 template <class T>
 ::mediapipe::Status TfLiteConverterCalculator::NormalizeImage(
    const ImageFrame& image_frame, bool zero_center, bool flip_vertically,
-    float* tensor_buffer) {
+    float* tensor_ptr) {
  const int height = image_frame.Height();
  const int width = image_frame.Width();
  const int channels = image_frame.NumberOfChannels();
@@ -651,7 +665,7 @@ template <class T>
        (flip_vertically ? height - 1 - i : i) * image_frame.WidthStep());
    for (int j = 0; j < width; ++j) {
      for (int c = 0; c < channels_preserved; ++c) {
-        *tensor_buffer++ = *image_ptr++ / div - sub;
+        *tensor_ptr++ = *image_ptr++ / div - sub;
      }
      image_ptr += channels_ignored;
    }
@@ -661,14 +675,14 @@ template <class T>
 }

 ::mediapipe::Status TfLiteConverterCalculator::CopyMatrixToTensor(
-    const Matrix& matrix, float* tensor_buffer) {
+    const Matrix& matrix, float* tensor_ptr) {
  if (row_major_matrix_) {
-    auto matrix_map = Eigen::Map<RowMajorMatrixXf>(tensor_buffer, matrix.rows(),
-                                                   matrix.cols());
+    auto matrix_map =
+        Eigen::Map<RowMajorMatrixXf>(tensor_ptr, matrix.rows(), matrix.cols());
    matrix_map = matrix;
  } else {
-    auto matrix_map = Eigen::Map<ColMajorMatrixXf>(tensor_buffer, matrix.rows(),
-                                                   matrix.cols());
+    auto matrix_map =
+        Eigen::Map<ColMajorMatrixXf>(tensor_ptr, matrix.rows(), matrix.cols());
    matrix_map = matrix;
  }

@@ -17,10 +17,16 @@
 #include <string>
 #include <vector>

+#include "absl/memory/memory.h"
 #include "mediapipe/calculators/tflite/tflite_inference_calculator.pb.h"
 #include "mediapipe/calculators/tflite/util.h"
 #include "mediapipe/framework/calculator_framework.h"
 #include "mediapipe/framework/port/ret_check.h"
+
+#if !defined(__EMSCRIPTEN__) || defined(__EMSCRIPTEN_PTHREADS__)
+#include "mediapipe/util/cpu_util.h"
+#endif  // !__EMSCRIPTEN__ || __EMSCRIPTEN_PTHREADS__
+
 #include "mediapipe/util/resource_util.h"
 #include "tensorflow/lite/error_reporter.h"
 #include "tensorflow/lite/interpreter.h"
@@ -30,6 +36,7 @@
 #if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
 #include "mediapipe/gpu/gl_calculator_helper.h"
 #include "mediapipe/gpu/gpu_buffer.h"
+#include "mediapipe/util/tflite/tflite_gpu_runner.h"
 #include "tensorflow/lite/delegates/gpu/common/shape.h"
 #include "tensorflow/lite/delegates/gpu/gl/gl_buffer.h"
 #include "tensorflow/lite/delegates/gpu/gl/gl_program.h"
@@ -50,7 +57,7 @@
 #include "tensorflow/lite/delegates/gpu/metal_delegate.h"
 #include "tensorflow/lite/delegates/gpu/metal_delegate_internal.h"
 #endif  // iOS
-
+#include "tensorflow/lite/delegates/xnnpack/xnnpack_delegate.h"
 #if defined(MEDIAPIPE_ANDROID)
 #include "tensorflow/lite/delegates/nnapi/nnapi_delegate.h"
 #endif  // ANDROID
@@ -69,6 +76,9 @@ typedef id<MTLBuffer> GpuTensor;

 // Round up n to next multiple of m.
 size_t RoundUp(size_t n, size_t m) { return ((n + m - 1) / m) * m; }  // NOLINT
+
+constexpr char kTensorsTag[] = "TENSORS";
+constexpr char kTensorsGpuTag[] = "TENSORS_GPU";
 }  // namespace

 #if defined(MEDIAPIPE_EDGE_TPU)
@@ -113,6 +123,23 @@ struct GPUData {
 };
 #endif

+// Returns number of threads to configure XNNPACK delegate with.
+// (Equal to user provided value if specified.  Otherwise, it returns number of
+// high cores (hard-coded to 1 for Emscripten without Threads extension))
+int GetXnnpackNumThreads(
+    const mediapipe::TfLiteInferenceCalculatorOptions& opts) {
+  static constexpr int kDefaultNumThreads = -1;
+  if (opts.has_delegate() && opts.delegate().has_xnnpack() &&
+      opts.delegate().xnnpack().num_threads() != kDefaultNumThreads) {
+    return opts.delegate().xnnpack().num_threads();
+  }
+#if !defined(__EMSCRIPTEN__) || defined(__EMSCRIPTEN_PTHREADS__)
+  return InferHigherCoreIds().size();
+#else
+  return 1;
+#endif  // !__EMSCRIPTEN__ || __EMSCRIPTEN_PTHREADS__
+}
+
 // Calculator Header Section

 // Runs inference on the provided input TFLite tensors and TFLite model.
@@ -139,6 +166,9 @@ struct GPUData {
 // Input side packet:
 //  CUSTOM_OP_RESOLVER (optional) - Use a custom op resolver,
 //                                  instead of the builtin one.
+//  MODEL (optional) - Use to specify TfLite model
+//                     (std::unique_ptr<tflite::FlatBufferModel,
+//                       std::function<void(tflite::FlatBufferModel*)>>)
 //
 // Example use:
 // node {
@@ -153,6 +183,20 @@ struct GPUData {
 //   }
 // }
 //
+// or
+//
+// node {
+//   calculator: "TfLiteInferenceCalculator"
+//   input_stream: "TENSORS:tensor_image"
+//   input_side_packet: "MODEL:model"
+//   output_stream: "TENSORS:tensors"
+//   options: {
+//     [mediapipe.TfLiteInferenceCalculatorOptions.ext] {
+//       delegate { gpu {} }
+//     }
+//   }
+// }
+//
 // IMPORTANT Notes:
 //  Tensors are assumed to be ordered correctly (sequentially added to model).
 //  Input tensors are assumed to be of the correct size and already normalized.
@@ -165,6 +209,9 @@ class TfLiteInferenceCalculator : public CalculatorBase {
 public:
  using TfLiteDelegatePtr =
      std::unique_ptr<TfLiteDelegate, std::function<void(TfLiteDelegate*)>>;
+  using TfLiteModelPtr =
+      std::unique_ptr<tflite::FlatBufferModel,
+                      std::function<void(tflite::FlatBufferModel*)>>;

  static ::mediapipe::Status GetContract(CalculatorContract* cc);

@@ -173,18 +220,20 @@ class TfLiteInferenceCalculator : public CalculatorBase {
  ::mediapipe::Status Close(CalculatorContext* cc) override;

 private:
-  ::mediapipe::Status LoadOptions(CalculatorContext* cc);
  ::mediapipe::Status LoadModel(CalculatorContext* cc);
+  ::mediapipe::StatusOr<Packet> GetModelAsPacket(const CalculatorContext& cc);
  ::mediapipe::Status LoadDelegate(CalculatorContext* cc);
+  ::mediapipe::Status InitTFLiteGPURunner();

+  Packet model_packet_;
  std::unique_ptr<tflite::Interpreter> interpreter_;
-  std::unique_ptr<tflite::FlatBufferModel> model_;
  TfLiteDelegatePtr delegate_;

 #if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
  mediapipe::GlCalculatorHelper gpu_helper_;
  std::vector<std::unique_ptr<GPUData>> gpu_data_in_;
  std::vector<std::unique_ptr<GPUData>> gpu_data_out_;
+  std::unique_ptr<tflite::gpu::TFLiteGPURunner> tflite_gpu_runner_;
 #elif defined(MEDIAPIPE_IOS)
  MPPMetalHelper* gpu_helper_ = nullptr;
  std::vector<std::unique_ptr<GPUData>> gpu_data_in_;
@@ -198,11 +247,12 @@ class TfLiteInferenceCalculator : public CalculatorBase {
      edgetpu::EdgeTpuManager::GetSingleton()->OpenDevice();
 #endif

-  std::string model_path_ = "";
  bool gpu_inference_ = false;
  bool gpu_input_ = false;
  bool gpu_output_ = false;
  bool use_quantized_tensors_ = false;
+
+  bool use_advanced_gpu_api_ = false;
 };
 REGISTER_CALCULATOR(TfLiteInferenceCalculator);

@@ -210,36 +260,40 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);

 ::mediapipe::Status TfLiteInferenceCalculator::GetContract(
    CalculatorContract* cc) {
-  RET_CHECK(cc->Inputs().HasTag("TENSORS") ^
-            cc->Inputs().HasTag("TENSORS_GPU"));
-  RET_CHECK(cc->Outputs().HasTag("TENSORS") ^
-            cc->Outputs().HasTag("TENSORS_GPU"));
+  RET_CHECK(cc->Inputs().HasTag(kTensorsTag) ^
+            cc->Inputs().HasTag(kTensorsGpuTag));
+  RET_CHECK(cc->Outputs().HasTag(kTensorsTag) ^
+            cc->Outputs().HasTag(kTensorsGpuTag));

  const auto& options =
      cc->Options<::mediapipe::TfLiteInferenceCalculatorOptions>();
+  RET_CHECK(!options.model_path().empty() ^
+            cc->InputSidePackets().HasTag("MODEL"))
+      << "Either model as side packet or model path in options is required.";
+
  bool use_gpu =
      options.has_delegate() ? options.delegate().has_gpu() : options.use_gpu();

-  if (cc->Inputs().HasTag("TENSORS"))
-    cc->Inputs().Tag("TENSORS").Set<std::vector<TfLiteTensor>>();
+  if (cc->Inputs().HasTag(kTensorsTag))
+    cc->Inputs().Tag(kTensorsTag).Set<std::vector<TfLiteTensor>>();
 #if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__)
-  if (cc->Inputs().HasTag("TENSORS_GPU")) {
+  if (cc->Inputs().HasTag(kTensorsGpuTag)) {
    RET_CHECK(!options.has_delegate() || options.delegate().has_gpu())
        << "GPU input is compatible with GPU delegate only.";

-    cc->Inputs().Tag("TENSORS_GPU").Set<std::vector<GpuTensor>>();
+    cc->Inputs().Tag(kTensorsGpuTag).Set<std::vector<GpuTensor>>();
    use_gpu |= true;
  }
 #endif  //  !MEDIAPIPE_DISABLE_GPU

-  if (cc->Outputs().HasTag("TENSORS"))
-    cc->Outputs().Tag("TENSORS").Set<std::vector<TfLiteTensor>>();
+  if (cc->Outputs().HasTag(kTensorsTag))
+    cc->Outputs().Tag(kTensorsTag).Set<std::vector<TfLiteTensor>>();
 #if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__)
-  if (cc->Outputs().HasTag("TENSORS_GPU")) {
+  if (cc->Outputs().HasTag(kTensorsGpuTag)) {
    RET_CHECK(!options.has_delegate() || options.delegate().has_gpu())
        << "GPU output is compatible with GPU delegate only.";

-    cc->Outputs().Tag("TENSORS_GPU").Set<std::vector<GpuTensor>>();
+    cc->Outputs().Tag(kTensorsGpuTag).Set<std::vector<GpuTensor>>();
    use_gpu |= true;
  }
 #endif  //  !MEDIAPIPE_DISABLE_GPU
@@ -249,6 +303,9 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
        .Tag("CUSTOM_OP_RESOLVER")
        .Set<tflite::ops::builtin::BuiltinOpResolver>();
  }
+  if (cc->InputSidePackets().HasTag("MODEL")) {
+    cc->InputSidePackets().Tag("MODEL").Set<TfLiteModelPtr>();
+  }

  if (use_gpu) {
 #if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
@@ -267,29 +324,35 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
 ::mediapipe::Status TfLiteInferenceCalculator::Open(CalculatorContext* cc) {
  cc->SetOffset(TimestampDiff(0));

-  MP_RETURN_IF_ERROR(LoadOptions(cc));
+  const auto& options =
+      cc->Options<::mediapipe::TfLiteInferenceCalculatorOptions>();
+  gpu_inference_ = options.use_gpu();

-  if (cc->Inputs().HasTag("TENSORS_GPU")) {
+  if (cc->Inputs().HasTag(kTensorsGpuTag)) {
 #if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__)
    gpu_input_ = true;
    gpu_inference_ = true;  // Inference must be on GPU also.
 #else
-    RET_CHECK(!cc->Inputs().HasTag("TENSORS_GPU"))
+    RET_CHECK(!cc->Inputs().HasTag(kTensorsGpuTag))
        << "GPU processing not enabled.";
 #endif  //  !MEDIAPIPE_DISABLE_GPU
  }

-  if (cc->Outputs().HasTag("TENSORS_GPU")) {
+  if (cc->Outputs().HasTag(kTensorsGpuTag)) {
 #if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__)
    gpu_output_ = true;
-    RET_CHECK(cc->Inputs().HasTag("TENSORS_GPU"))
+    RET_CHECK(cc->Inputs().HasTag(kTensorsGpuTag))
        << "GPU output must also have GPU Input.";
 #else
-    RET_CHECK(!cc->Inputs().HasTag("TENSORS_GPU"))
+    RET_CHECK(!cc->Inputs().HasTag(kTensorsGpuTag))
        << "GPU processing not enabled.";
 #endif  //  !MEDIAPIPE_DISABLE_GPU
  }

+  const auto& calculator_opts =
+      cc->Options<mediapipe::TfLiteInferenceCalculatorOptions>();
+  use_advanced_gpu_api_ = false;
+
  MP_RETURN_IF_ERROR(LoadModel(cc));

  if (gpu_inference_) {
@@ -301,8 +364,12 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
 #endif

 #if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-    MP_RETURN_IF_ERROR(gpu_helper_.RunInGlContext(
-        [this, &cc]() -> ::mediapipe::Status { return LoadDelegate(cc); }));
+    MP_RETURN_IF_ERROR(
+        gpu_helper_.RunInGlContext([this, &cc]() -> ::mediapipe::Status {
+          return use_advanced_gpu_api_ ? InitTFLiteGPURunner()
+                                       : LoadDelegate(cc);
+        }));
+    if (use_advanced_gpu_api_) return ::mediapipe::OkStatus();
 #else
    MP_RETURN_IF_ERROR(LoadDelegate(cc));
 #endif
@@ -314,13 +381,51 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
  return ::mediapipe::OkStatus();
 }

+::mediapipe::Status TfLiteInferenceCalculator::InitTFLiteGPURunner() {
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
+  // Create and bind OpenGL buffers for outputs.
+  // These buffers are created onve and later their ids are jut passed to the
+  // calculator outputs.
+
+  gpu_data_out_.resize(tflite_gpu_runner_->outputs_size());
+  for (int i = 0; i < tflite_gpu_runner_->outputs_size(); ++i) {
+    gpu_data_out_[i] = absl::make_unique<GPUData>();
+    ASSIGN_OR_RETURN(gpu_data_out_[i]->elements,
+                     tflite_gpu_runner_->GetOutputElements(i));
+    // Create and bind input buffer.
+    RET_CHECK_CALL(::tflite::gpu::gl::CreateReadWriteShaderStorageBuffer<float>(
+        gpu_data_out_[i]->elements, &gpu_data_out_[i]->buffer));
+  }
+  RET_CHECK_CALL(tflite_gpu_runner_->Build());
+#endif
+  return ::mediapipe::OkStatus();
+}
+
 ::mediapipe::Status TfLiteInferenceCalculator::Process(CalculatorContext* cc) {
  // 1. Receive pre-processed tensor inputs.
-  if (gpu_input_) {
-    // Read GPU input into SSBO.
+  if (use_advanced_gpu_api_) {
 #if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
    const auto& input_tensors =
        cc->Inputs().Tag("TENSORS_GPU").Get<std::vector<GpuTensor>>();
+    RET_CHECK(input_tensors.empty());
+    MP_RETURN_IF_ERROR(gpu_helper_.RunInGlContext(
+        [this, &input_tensors]() -> ::mediapipe::Status {
+          for (int i = 0; i < input_tensors.size(); ++i) {
+            MP_RETURN_IF_ERROR(tflite_gpu_runner_->BindSSBOToInputTensor(
+                input_tensors[i].id(), i));
+          }
+          for (int i = 0; i < gpu_data_out_.size(); ++i) {
+            MP_RETURN_IF_ERROR(tflite_gpu_runner_->BindSSBOToOutputTensor(
+                gpu_data_out_[i]->buffer.id(), i));
+          }
+          return ::mediapipe::OkStatus();
+        }));
+#endif
+  } else if (gpu_input_) {
+    // Read GPU input into SSBO.
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
+    const auto& input_tensors =
+        cc->Inputs().Tag(kTensorsGpuTag).Get<std::vector<GpuTensor>>();
    RET_CHECK_GT(input_tensors.size(), 0);
    MP_RETURN_IF_ERROR(gpu_helper_.RunInGlContext(
        [this, &input_tensors]() -> ::mediapipe::Status {
@@ -335,7 +440,7 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
        }));
 #elif defined(MEDIAPIPE_IOS)
    const auto& input_tensors =
-        cc->Inputs().Tag("TENSORS_GPU").Get<std::vector<GpuTensor>>();
+        cc->Inputs().Tag(kTensorsGpuTag).Get<std::vector<GpuTensor>>();
    RET_CHECK_GT(input_tensors.size(), 0);
    // Explicit copy input with conversion float 32 bits to 16 bits.
    gpu_data_in_.resize(input_tensors.size());
@@ -362,7 +467,7 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
  } else {
    // Read CPU input into tensors.
    const auto& input_tensors =
-        cc->Inputs().Tag("TENSORS").Get<std::vector<TfLiteTensor>>();
+        cc->Inputs().Tag(kTensorsTag).Get<std::vector<TfLiteTensor>>();
    RET_CHECK_GT(input_tensors.size(), 0);
    for (int i = 0; i < input_tensors.size(); ++i) {
      const TfLiteTensor* input_tensor = &input_tensors[i];
@@ -386,7 +491,11 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
 #if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
    MP_RETURN_IF_ERROR(
        gpu_helper_.RunInGlContext([this]() -> ::mediapipe::Status {
-          RET_CHECK_EQ(interpreter_->Invoke(), kTfLiteOk);
+          if (use_advanced_gpu_api_) {
+            RET_CHECK(tflite_gpu_runner_->Invoke().ok());
+          } else {
+            RET_CHECK_EQ(interpreter_->Invoke(), kTfLiteOk);
+          }
          return ::mediapipe::OkStatus();
        }));
 #elif defined(MEDIAPIPE_IOS)
@@ -397,7 +506,18 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
  }

  // 3. Output processed tensors.
-  if (gpu_output_) {
+  if (use_advanced_gpu_api_) {
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
+    auto output_tensors = absl::make_unique<std::vector<GpuTensor>>();
+    output_tensors->resize(gpu_data_out_.size());
+    for (int i = 0; i < gpu_data_out_.size(); ++i) {
+      output_tensors->at(i) = gpu_data_out_[0]->buffer.MakeRef();
+    }
+    cc->Outputs()
+        .Tag("TENSORS_GPU")
+        .Add(output_tensors.release(), cc->InputTimestamp());
+#endif
+  } else if (gpu_output_) {
 #if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
    // Output result tensors (GPU).
    auto output_tensors = absl::make_unique<std::vector<GpuTensor>>();
@@ -413,7 +533,7 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
          return ::mediapipe::OkStatus();
        }));
    cc->Outputs()
-        .Tag("TENSORS_GPU")
+        .Tag(kTensorsGpuTag)
        .Add(output_tensors.release(), cc->InputTimestamp());
 #elif defined(MEDIAPIPE_IOS)
    // Output result tensors (GPU).
@@ -437,7 +557,7 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
    [convert_command endEncoding];
    [command_buffer commit];
    cc->Outputs()
-        .Tag("TENSORS_GPU")
+        .Tag(kTensorsGpuTag)
        .Add(output_tensors.release(), cc->InputTimestamp());
 #else
    RET_CHECK_FAIL() << "GPU processing not enabled.";
@@ -450,8 +570,9 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
      TfLiteTensor* tensor = interpreter_->tensor(tensor_indexes[i]);
      output_tensors->emplace_back(*tensor);
    }
-    cc->Outputs().Tag("TENSORS").Add(output_tensors.release(),
-                                     cc->InputTimestamp());
+    cc->Outputs()
+        .Tag(kTensorsTag)
+        .Add(output_tensors.release(), cc->InputTimestamp());
  }

  return ::mediapipe::OkStatus();
@@ -462,6 +583,7 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
    if (gpu_inference_) {
 #if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
      MP_RETURN_IF_ERROR(gpu_helper_.RunInGlContext([this]() -> Status {
+        interpreter_ = nullptr;
        delegate_ = nullptr;
        for (int i = 0; i < gpu_data_in_.size(); ++i) {
          gpu_data_in_[i].reset();
@@ -472,6 +594,7 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
        return ::mediapipe::OkStatus();
      }));
 #elif defined(MEDIAPIPE_IOS)
+      interpreter_ = nullptr;
      delegate_ = nullptr;
      for (int i = 0; i < gpu_data_in_.size(); ++i) {
        gpu_data_in_[i].reset();
@@ -481,6 +604,7 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
      }
 #endif
    } else {
+      interpreter_ = nullptr;
      delegate_ = nullptr;
    }
  }
@@ -492,34 +616,10 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);

 // Calculator Auxiliary Section

-::mediapipe::Status TfLiteInferenceCalculator::LoadOptions(
-    CalculatorContext* cc) {
-  // Get calculator options specified in the graph.
-  const auto& options =
-      cc->Options<::mediapipe::TfLiteInferenceCalculatorOptions>();
-
-  // Get model name.
-  if (!options.model_path().empty()) {
-    std::string model_path = options.model_path();
-
-    ASSIGN_OR_RETURN(model_path_, mediapipe::PathToResourceAsFile(model_path));
-  } else {
-    LOG(ERROR) << "Must specify path to TFLite model.";
-    return ::mediapipe::Status(::mediapipe::StatusCode::kNotFound,
-                               "Must specify path to TFLite model.");
-  }
-
-  // Get execution modes.
-  gpu_inference_ =
-      options.has_delegate() ? options.delegate().has_gpu() : options.use_gpu();
-
-  return ::mediapipe::OkStatus();
-}
-
 ::mediapipe::Status TfLiteInferenceCalculator::LoadModel(
    CalculatorContext* cc) {
-  model_ = tflite::FlatBufferModel::BuildFromFile(model_path_.c_str());
-  RET_CHECK(model_);
+  ASSIGN_OR_RETURN(model_packet_, GetModelAsPacket(*cc));
+  const auto& model = *model_packet_.Get<TfLiteModelPtr>();

  tflite::ops::builtin::BuiltinOpResolver op_resolver;
  if (cc->InputSidePackets().HasTag("CUSTOM_OP_RESOLVER")) {
@@ -527,11 +627,25 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
                      .Tag("CUSTOM_OP_RESOLVER")
                      .Get<tflite::ops::builtin::BuiltinOpResolver>();
  }
+
+#if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
+  if (use_advanced_gpu_api_) {
+    tflite::gpu::InferenceOptions options;
+    options.priority1 = tflite::gpu::InferencePriority::MIN_LATENCY;
+    options.priority2 = tflite::gpu::InferencePriority::AUTO;
+    options.priority3 = tflite::gpu::InferencePriority::AUTO;
+    options.usage = tflite::gpu::InferenceUsage::SUSTAINED_SPEED;
+    tflite_gpu_runner_ =
+        std::make_unique<tflite::gpu::TFLiteGPURunner>(options);
+    return tflite_gpu_runner_->InitializeWithModel(model);
+  }
+#endif
+
 #if defined(MEDIAPIPE_EDGE_TPU)
  interpreter_ =
-      BuildEdgeTpuInterpreter(*model_, &op_resolver, edgetpu_context_.get());
+      BuildEdgeTpuInterpreter(model, &op_resolver, edgetpu_context_.get());
 #else
-  tflite::InterpreterBuilder(*model_, op_resolver)(&interpreter_);
+  tflite::InterpreterBuilder(model, op_resolver)(&interpreter_);
 #endif  // MEDIAPIPE_EDGE_TPU

  RET_CHECK(interpreter_);
@@ -557,6 +671,28 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
  return ::mediapipe::OkStatus();
 }

+::mediapipe::StatusOr<Packet> TfLiteInferenceCalculator::GetModelAsPacket(
+    const CalculatorContext& cc) {
+  const auto& options =
+      cc.Options<mediapipe::TfLiteInferenceCalculatorOptions>();
+  if (!options.model_path().empty()) {
+    std::string model_path = options.model_path();
+
+    ASSIGN_OR_RETURN(model_path, mediapipe::PathToResourceAsFile(model_path));
+
+    auto model = tflite::FlatBufferModel::BuildFromFile(model_path.c_str());
+    RET_CHECK(model) << "Failed to load model from path.";
+    return MakePacket<TfLiteModelPtr>(TfLiteModelPtr(
+        model.release(), [](tflite::FlatBufferModel* model) { delete model; }));
+  }
+  if (cc.InputSidePackets().HasTag("MODEL")) {
+    return cc.InputSidePackets().Tag("MODEL");
+  }
+  return ::mediapipe::Status(
+      ::mediapipe::StatusCode::kNotFound,
+      "Must specify TFLite model as path or loaded model.");
+}
+
 ::mediapipe::Status TfLiteInferenceCalculator::LoadDelegate(
    CalculatorContext* cc) {
  const auto& calculator_opts =
@@ -587,6 +723,22 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
    }
 #endif  // MEDIAPIPE_ANDROID

+#if defined(__EMSCRIPTEN__)
+    const bool xnnpack_requested = true;
+#else
+    const bool xnnpack_requested = calculator_opts.has_delegate() &&
+                                   calculator_opts.delegate().has_xnnpack();
+#endif  // __EMSCRIPTEN__
+
+    if (xnnpack_requested) {
+      TfLiteXNNPackDelegateOptions xnnpack_opts{};
+      xnnpack_opts.num_threads = GetXnnpackNumThreads(calculator_opts);
+      delegate_ = TfLiteDelegatePtr(TfLiteXNNPackDelegateCreate(&xnnpack_opts),
+                                    &TfLiteXNNPackDelegateDelete);
+      RET_CHECK_EQ(interpreter_->ModifyGraphWithDelegate(delegate_.get()),
+                   kTfLiteOk);
+    }
+
    // Return, no need for GPU delegate below.
    return ::mediapipe::OkStatus();
  }
@@ -608,7 +760,7 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
    const auto& input_indices = interpreter_->inputs();
    gpu_data_in_.resize(input_indices.size());
    for (int i = 0; i < input_indices.size(); ++i) {
-      const TfLiteTensor* tensor = interpreter_->tensor(input_indices[0]);
+      const TfLiteTensor* tensor = interpreter_->tensor(input_indices[i]);
      gpu_data_in_[i] = absl::make_unique<GPUData>();
      gpu_data_in_[i]->elements = 1;
      for (int d = 0; d < tensor->dims->size; ++d) {
@@ -42,14 +42,24 @@ message TfLiteInferenceCalculatorOptions {
    message TfLite {}
    // Delegate to run GPU inference depending on the device.
    // (Can use OpenGl, OpenCl, Metal depending on the device.)
-    message Gpu {}
+    message Gpu {
+      // Experimental, Android/Linux only. Use TFLite GPU delegate API2 for
+      // the NN inference.
+      optional bool use_advanced_gpu_api = 1 [default = false];
+    }
    // Android only.
    message Nnapi {}
+    message Xnnpack {
+      // Number of threads for XNNPACK delegate. (By default, calculator tries
+      // to choose optimal number of threads depending on the device.)
+      optional int32 num_threads = 1 [default = -1];
+    }

    oneof delegate {
      TfLite tflite = 1;
      Gpu gpu = 2;
      Nnapi nnapi = 3;
+      Xnnpack xnnpack = 4;
    }
  }

@@ -41,7 +41,7 @@ namespace mediapipe {

 using ::tflite::Interpreter;

-void DoSmokeTest(absl::string_view delegate) {
+void DoSmokeTest(const std::string& graph_proto) {
  const int width = 8;
  const int height = 8;
  const int channels = 3;
@@ -69,24 +69,9 @@ void DoSmokeTest(absl::string_view delegate) {
  auto input_vec = absl::make_unique<std::vector<TfLiteTensor>>();
  input_vec->emplace_back(*tensor);

-  std::string graph_proto = R"(
-    input_stream: "tensor_in"
-    node {
-      calculator: "TfLiteInferenceCalculator"
-      input_stream: "TENSORS:tensor_in"
-      output_stream: "TENSORS:tensor_out"
-      options {
-        [mediapipe.TfLiteInferenceCalculatorOptions.ext] {
-          model_path: "mediapipe/calculators/tflite/testdata/add.bin"
-          $delegate
-        }
-      }
-    }
-  )";
-  ASSERT_EQ(absl::StrReplaceAll({{"$delegate", delegate}}, &graph_proto), 1);
  // Prepare single calculator graph to and wait for packets.
  CalculatorGraphConfig graph_config =
-      ::mediapipe::ParseTextProtoOrDie<CalculatorGraphConfig>(graph_proto);
+      ParseTextProtoOrDie<CalculatorGraphConfig>(graph_proto);
  std::vector<Packet> output_packets;
  tool::AddVectorSink("tensor_out", &graph_config, &output_packets);
  CalculatorGraph graph(graph_config);
@@ -119,8 +104,70 @@ void DoSmokeTest(absl::string_view delegate) {

 // Tests a simple add model that adds an input tensor to itself.
 TEST(TfLiteInferenceCalculatorTest, SmokeTest) {
-  DoSmokeTest(/*delegate=*/"");
-  DoSmokeTest(/*delegate=*/"delegate { tflite {} }");
+  std::string graph_proto = R"(
+    input_stream: "tensor_in"
+    node {
+      calculator: "TfLiteInferenceCalculator"
+      input_stream: "TENSORS:tensor_in"
+      output_stream: "TENSORS:tensor_out"
+      options {
+        [mediapipe.TfLiteInferenceCalculatorOptions.ext] {
+          model_path: "mediapipe/calculators/tflite/testdata/add.bin"
+          $delegate
+        }
+      }
+    }
+  )";
+  DoSmokeTest(
+      /*graph_proto=*/absl::StrReplaceAll(graph_proto, {{"$delegate", ""}}));
+  DoSmokeTest(/*graph_proto=*/absl::StrReplaceAll(
+      graph_proto, {{"$delegate", "delegate { tflite {} }"}}));
+  DoSmokeTest(/*graph_proto=*/absl::StrReplaceAll(
+      graph_proto, {{"$delegate", "delegate { xnnpack {} }"}}));
+  DoSmokeTest(/*graph_proto=*/absl::StrReplaceAll(
+      graph_proto,
+      {{"$delegate", "delegate { xnnpack { num_threads: 10 } }"}}));
+}
+
+TEST(TfLiteInferenceCalculatorTest, SmokeTest_ModelAsInputSidePacket) {
+  std::string graph_proto = R"(
+    input_stream: "tensor_in"
+
+    node {
+      calculator: "ConstantSidePacketCalculator"
+      output_side_packet: "PACKET:model_path"
+      options: {
+        [mediapipe.ConstantSidePacketCalculatorOptions.ext]: {
+          packet { string_value: "mediapipe/calculators/tflite/testdata/add.bin" }
+        }
+      }
+    }
+
+    node {
+      calculator: "LocalFileContentsCalculator"
+      input_side_packet: "FILE_PATH:model_path"
+      output_side_packet: "CONTENTS:model_blob"
+    }
+
+    node {
+      calculator: "TfLiteModelCalculator"
+      input_side_packet: "MODEL_BLOB:model_blob"
+      output_side_packet: "MODEL:model"
+    }
+
+    node {
+      calculator: "TfLiteInferenceCalculator"
+      input_stream: "TENSORS:tensor_in"
+      output_stream: "TENSORS:tensor_out"
+      input_side_packet: "MODEL:model"
+      options {
+        [mediapipe.TfLiteInferenceCalculatorOptions.ext] {
+          use_gpu: false
+        }
+      }
+    }
+  )";
+  DoSmokeTest(graph_proto);
 }

 }  // namespace mediapipe
@@ -0,0 +1,86 @@
+// Copyright 2020 The MediaPipe Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <functional>
+#include <memory>
+#include <string>
+
+#include "mediapipe/framework/calculator_framework.h"
+#include "mediapipe/framework/packet.h"
+#include "mediapipe/framework/port/ret_check.h"
+#include "tensorflow/lite/model.h"
+
+namespace mediapipe {
+
+// Loads TfLite model from model blob specified as input side packet and outputs
+// corresponding side packet.
+//
+// Input side packets:
+//   MODEL_BLOB - TfLite model blob/file-contents (std::string). You can read
+//                model blob from file (using whatever APIs you have) and pass
+//                it to the graph as input side packet or you can use some of
+//                calculators like LocalFileContentsCalculator to get model
+//                blob and use it as input here.
+//
+// Output side packets:
+//   MODEL - TfLite model. (std::unique_ptr<tflite::FlatBufferModel,
+//           std::function<void(tflite::FlatBufferModel*)>>)
+//
+// Example use:
+//
+// node {
+//   calculator: "TfLiteModelCalculator"
+//   input_side_packet: "MODEL_BLOB:model_blob"
+//   output_side_packet: "MODEL:model"
+// }
+//
+class TfLiteModelCalculator : public CalculatorBase {
+ public:
+  using TfLiteModelPtr =
+      std::unique_ptr<tflite::FlatBufferModel,
+                      std::function<void(tflite::FlatBufferModel*)>>;
+
+  static ::mediapipe::Status GetContract(CalculatorContract* cc) {
+    cc->InputSidePackets().Tag("MODEL_BLOB").Set<std::string>();
+    cc->OutputSidePackets().Tag("MODEL").Set<TfLiteModelPtr>();
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Open(CalculatorContext* cc) override {
+    const Packet& model_packet = cc->InputSidePackets().Tag("MODEL_BLOB");
+    const std::string& model_blob = model_packet.Get<std::string>();
+    std::unique_ptr<tflite::FlatBufferModel> model =
+        tflite::FlatBufferModel::BuildFromBuffer(model_blob.data(),
+                                                 model_blob.size());
+    RET_CHECK(model) << "Failed to load TfLite model from blob.";
+
+    cc->OutputSidePackets().Tag("MODEL").Set(
+        MakePacket<TfLiteModelPtr>(TfLiteModelPtr(
+            model.release(), [model_packet](tflite::FlatBufferModel* model) {
+              // Keeping model_packet in order to keep underlying model blob
+              // which can be released only after TfLite model is not needed
+              // anymore (deleted).
+              delete model;
+            })));
+
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Process(CalculatorContext* cc) override {
+    return ::mediapipe::OkStatus();
+  }
+};
+REGISTER_CALCULATOR(TfLiteModelCalculator);
+
+}  // namespace mediapipe
@@ -0,0 +1,88 @@
+// Copyright 2020 The MediaPipe Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <memory>
+
+#include "mediapipe/framework/calculator_framework.h"
+#include "mediapipe/framework/calculator_runner.h"
+#include "mediapipe/framework/port/gmock.h"
+#include "mediapipe/framework/port/gtest.h"
+#include "mediapipe/framework/port/parse_text_proto.h"
+#include "mediapipe/framework/port/status_matchers.h"  // NOLINT
+#include "tensorflow/lite/model.h"
+
+namespace mediapipe {
+
+TEST(TfLiteModelCalculatorTest, SmokeTest) {
+  // Prepare single calculator graph to and wait for packets.
+  CalculatorGraphConfig graph_config = ParseTextProtoOrDie<
+      CalculatorGraphConfig>(
+      R"(
+        node {
+          calculator: "ConstantSidePacketCalculator"
+          output_side_packet: "PACKET:model_path"
+          options: {
+            [mediapipe.ConstantSidePacketCalculatorOptions.ext]: {
+              packet {
+                string_value: "mediapipe/calculators/tflite/testdata/add.bin"
+              }
+            }
+          }
+        }
+
+        node {
+          calculator: "LocalFileContentsCalculator"
+          input_side_packet: "FILE_PATH:model_path"
+          output_side_packet: "CONTENTS:model_blob"
+        }
+
+        node {
+          calculator: "TfLiteModelCalculator"
+          input_side_packet: "MODEL_BLOB:model_blob"
+          output_side_packet: "MODEL:model"
+        }
+      )");
+  CalculatorGraph graph(graph_config);
+  MP_ASSERT_OK(graph.StartRun({}));
+  MP_ASSERT_OK(graph.WaitUntilIdle());
+  auto status_or_packet = graph.GetOutputSidePacket("model");
+  MP_ASSERT_OK(status_or_packet);
+  auto model_packet = status_or_packet.ValueOrDie();
+  const auto& model = model_packet.Get<
+      std::unique_ptr<tflite::FlatBufferModel,
+                      std::function<void(tflite::FlatBufferModel*)>>>();
+
+  auto expected_model = tflite::FlatBufferModel::BuildFromFile(
+      "mediapipe/calculators/tflite/testdata/add.bin");
+
+  EXPECT_EQ(model->GetModel()->version(),
+            expected_model->GetModel()->version());
+  EXPECT_EQ(model->GetModel()->buffers()->size(),
+            expected_model->GetModel()->buffers()->size());
+  const int num_subgraphs = expected_model->GetModel()->subgraphs()->size();
+  EXPECT_EQ(model->GetModel()->subgraphs()->size(), num_subgraphs);
+  for (int i = 0; i < num_subgraphs; ++i) {
+    const auto* expected_subgraph =
+        expected_model->GetModel()->subgraphs()->Get(i);
+    const auto* subgraph = model->GetModel()->subgraphs()->Get(i);
+    const int num_tensors = expected_subgraph->tensors()->size();
+    EXPECT_EQ(subgraph->tensors()->size(), num_tensors);
+    for (int j = 0; j < num_tensors; ++j) {
+      EXPECT_EQ(subgraph->tensors()->Get(j)->name()->str(),
+                expected_subgraph->tensors()->Get(j)->name()->str());
+    }
+  }
+}
+
+}  // namespace mediapipe
@@ -129,22 +129,43 @@ REGISTER_CALCULATOR(TfLiteTensorsToClassificationCalculator);
    num_classes *= raw_score_tensor->dims->data[i];
  }

+  if (options_.binary_classification()) {
+    RET_CHECK_EQ(num_classes, 1);
+    // Number of classes for binary classification.
+    num_classes = 2;
+  }
  if (label_map_loaded_) {
    RET_CHECK_EQ(num_classes, label_map_.size());
  }
  const float* raw_scores = raw_score_tensor->data.f;

  auto classification_list = absl::make_unique<ClassificationList>();
-  for (int i = 0; i < num_classes; ++i) {
-    if (options_.has_min_score_threshold() &&
-        raw_scores[i] < options_.min_score_threshold()) {
-      continue;
-    }
-    Classification* classification = classification_list->add_classification();
-    classification->set_index(i);
-    classification->set_score(raw_scores[i]);
+  if (options_.binary_classification()) {
+    Classification* class_first = classification_list->add_classification();
+    Classification* class_second = classification_list->add_classification();
+    class_first->set_index(0);
+    class_second->set_index(1);
+    class_first->set_score(raw_scores[0]);
+    class_second->set_score(1. - raw_scores[0]);
+
    if (label_map_loaded_) {
-      classification->set_label(label_map_[i]);
+      class_first->set_label(label_map_[0]);
+      class_second->set_label(label_map_[1]);
+    }
+  } else {
+    for (int i = 0; i < num_classes; ++i) {
+      if (options_.has_min_score_threshold() &&
+          raw_scores[i] < options_.min_score_threshold()) {
+        continue;
+      }
+      Classification* classification =
+          classification_list->add_classification();
+      classification->set_index(i);
+      classification->set_score(raw_scores[i]);
+
+      if (label_map_loaded_) {
+        classification->set_label(label_map_[i]);
+      }
    }
  }

@@ -32,4 +32,10 @@ message TfLiteTensorsToClassificationCalculatorOptions {
  optional int32 top_k = 2;
  // Path to a label map file for getting the actual name of class ids.
  optional string label_map_path = 3;
+  // Whether the input is a single float for binary classification.
+  // When true, only a single float is expected in the input tensor and the
+  // label map, if provided, is expected to have exactly two labels.
+  // The single score(float) represent the probability of first label, and
+  // 1 - score is the probabilility of the second label.
+  optional bool binary_classification = 4;
 }
@@ -47,10 +47,11 @@
 #endif  // iOS

 namespace {
-
 constexpr int kNumInputTensorsWithAnchors = 3;
 constexpr int kNumCoordsPerBox = 4;

+constexpr char kTensorsTag[] = "TENSORS";
+constexpr char kTensorsGpuTag[] = "TENSORS_GPU";
 }  // namespace

 namespace mediapipe {
@@ -99,7 +100,7 @@ void ConvertAnchorsToRawValues(const std::vector<Anchor>& anchors,
                               int num_boxes, float* raw_anchors) {
  CHECK_EQ(anchors.size(), num_boxes);
  int box = 0;
-  for (auto anchor : anchors) {
+  for (const auto& anchor : anchors) {
    raw_anchors[box * kNumCoordsPerBox + 0] = anchor.y_center();
    raw_anchors[box * kNumCoordsPerBox + 1] = anchor.x_center();
    raw_anchors[box * kNumCoordsPerBox + 2] = anchor.h();
@@ -200,13 +201,13 @@ REGISTER_CALCULATOR(TfLiteTensorsToDetectionsCalculator);

  bool use_gpu = false;

-  if (cc->Inputs().HasTag("TENSORS")) {
-    cc->Inputs().Tag("TENSORS").Set<std::vector<TfLiteTensor>>();
+  if (cc->Inputs().HasTag(kTensorsTag)) {
+    cc->Inputs().Tag(kTensorsTag).Set<std::vector<TfLiteTensor>>();
  }

 #if !defined(MEDIAPIPE_DISABLE_GPU) && !defined(__EMSCRIPTEN__)
-  if (cc->Inputs().HasTag("TENSORS_GPU")) {
-    cc->Inputs().Tag("TENSORS_GPU").Set<std::vector<GpuTensor>>();
+  if (cc->Inputs().HasTag(kTensorsGpuTag)) {
+    cc->Inputs().Tag(kTensorsGpuTag).Set<std::vector<GpuTensor>>();
    use_gpu |= true;
  }
 #endif  //  !MEDIAPIPE_DISABLE_GPU
@@ -236,7 +237,7 @@ REGISTER_CALCULATOR(TfLiteTensorsToDetectionsCalculator);
    CalculatorContext* cc) {
  cc->SetOffset(TimestampDiff(0));

-  if (cc->Inputs().HasTag("TENSORS_GPU")) {
+  if (cc->Inputs().HasTag(kTensorsGpuTag)) {
    gpu_input_ = true;
 #if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
    MP_RETURN_IF_ERROR(gpu_helper_.Open(cc));
@@ -258,8 +259,8 @@ REGISTER_CALCULATOR(TfLiteTensorsToDetectionsCalculator);

 ::mediapipe::Status TfLiteTensorsToDetectionsCalculator::Process(
    CalculatorContext* cc) {
-  if ((!gpu_input_ && cc->Inputs().Tag("TENSORS").IsEmpty()) ||
-      (gpu_input_ && cc->Inputs().Tag("TENSORS_GPU").IsEmpty())) {
+  if ((!gpu_input_ && cc->Inputs().Tag(kTensorsTag).IsEmpty()) ||
+      (gpu_input_ && cc->Inputs().Tag(kTensorsGpuTag).IsEmpty())) {
    return ::mediapipe::OkStatus();
  }

@@ -284,7 +285,7 @@ REGISTER_CALCULATOR(TfLiteTensorsToDetectionsCalculator);
 ::mediapipe::Status TfLiteTensorsToDetectionsCalculator::ProcessCPU(
    CalculatorContext* cc, std::vector<Detection>* output_detections) {
  const auto& input_tensors =
-      cc->Inputs().Tag("TENSORS").Get<std::vector<TfLiteTensor>>();
+      cc->Inputs().Tag(kTensorsTag).Get<std::vector<TfLiteTensor>>();

  if (input_tensors.size() == 2 ||
      input_tensors.size() == kNumInputTensorsWithAnchors) {
@@ -402,7 +403,7 @@ REGISTER_CALCULATOR(TfLiteTensorsToDetectionsCalculator);
    CalculatorContext* cc, std::vector<Detection>* output_detections) {
 #if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
  const auto& input_tensors =
-      cc->Inputs().Tag("TENSORS_GPU").Get<std::vector<GpuTensor>>();
+      cc->Inputs().Tag(kTensorsGpuTag).Get<std::vector<GpuTensor>>();
  RET_CHECK_GE(input_tensors.size(), 2);

  MP_RETURN_IF_ERROR(gpu_helper_.RunInGlContext([this, &input_tensors, &cc,
@@ -466,7 +467,7 @@ REGISTER_CALCULATOR(TfLiteTensorsToDetectionsCalculator);
 #elif defined(MEDIAPIPE_IOS)

  const auto& input_tensors =
-      cc->Inputs().Tag("TENSORS_GPU").Get<std::vector<GpuTensor>>();
+      cc->Inputs().Tag(kTensorsGpuTag).Get<std::vector<GpuTensor>>();
  RET_CHECK_GE(input_tensors.size(), 2);

  // Copy inputs.
@@ -145,7 +145,7 @@ REGISTER_CALCULATOR(TfLiteTensorsToLandmarksCalculator);
          ? cc->InputSidePackets().Tag("FLIP_HORIZONTALLY").Get<bool>()
          : options_.flip_horizontally();

-  flip_horizontally_ =
+  flip_vertically_ =
      cc->InputSidePackets().HasTag("FLIP_VERTICALLY")
          ? cc->InputSidePackets().Tag("FLIP_VERTICALLY").Get<bool>()
          : options_.flip_vertically();
@@ -49,6 +49,16 @@ int NumGroups(const int size, const int group_size) {  // NOLINT
 float Clamp(float val, float min, float max) {
  return std::min(std::max(val, min), max);
 }
+
+constexpr char kTensorsTag[] = "TENSORS";
+constexpr char kTensorsGpuTag[] = "TENSORS_GPU";
+constexpr char kSizeImageTag[] = "REFERENCE_IMAGE";
+constexpr char kSizeImageGpuTag[] = "REFERENCE_IMAGE_GPU";
+constexpr char kMaskTag[] = "MASK";
+constexpr char kMaskGpuTag[] = "MASK_GPU";
+constexpr char kPrevMaskTag[] = "PREV_MASK";
+constexpr char kPrevMaskGpuTag[] = "PREV_MASK_GPU";
+
 }  // namespace

 namespace mediapipe {
@@ -148,39 +158,39 @@ REGISTER_CALCULATOR(TfLiteTensorsToSegmentationCalculator);
  bool use_gpu = false;

  // Inputs CPU.
-  if (cc->Inputs().HasTag("TENSORS")) {
-    cc->Inputs().Tag("TENSORS").Set<std::vector<TfLiteTensor>>();
+  if (cc->Inputs().HasTag(kTensorsTag)) {
+    cc->Inputs().Tag(kTensorsTag).Set<std::vector<TfLiteTensor>>();
  }
-  if (cc->Inputs().HasTag("PREV_MASK")) {
-    cc->Inputs().Tag("PREV_MASK").Set<ImageFrame>();
+  if (cc->Inputs().HasTag(kPrevMaskTag)) {
+    cc->Inputs().Tag(kPrevMaskTag).Set<ImageFrame>();
  }
-  if (cc->Inputs().HasTag("REFERENCE_IMAGE")) {
-    cc->Inputs().Tag("REFERENCE_IMAGE").Set<ImageFrame>();
+  if (cc->Inputs().HasTag(kSizeImageTag)) {
+    cc->Inputs().Tag(kSizeImageTag).Set<ImageFrame>();
  }

  // Inputs GPU.
 #if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-  if (cc->Inputs().HasTag("TENSORS_GPU")) {
-    cc->Inputs().Tag("TENSORS_GPU").Set<std::vector<GlBuffer>>();
+  if (cc->Inputs().HasTag(kTensorsGpuTag)) {
+    cc->Inputs().Tag(kTensorsGpuTag).Set<std::vector<GlBuffer>>();
    use_gpu |= true;
  }
-  if (cc->Inputs().HasTag("PREV_MASK_GPU")) {
-    cc->Inputs().Tag("PREV_MASK_GPU").Set<mediapipe::GpuBuffer>();
+  if (cc->Inputs().HasTag(kPrevMaskGpuTag)) {
+    cc->Inputs().Tag(kPrevMaskGpuTag).Set<mediapipe::GpuBuffer>();
    use_gpu |= true;
  }
-  if (cc->Inputs().HasTag("REFERENCE_IMAGE_GPU")) {
-    cc->Inputs().Tag("REFERENCE_IMAGE_GPU").Set<mediapipe::GpuBuffer>();
+  if (cc->Inputs().HasTag(kSizeImageGpuTag)) {
+    cc->Inputs().Tag(kSizeImageGpuTag).Set<mediapipe::GpuBuffer>();
    use_gpu |= true;
  }
 #endif  //  !MEDIAPIPE_DISABLE_GPU

  // Outputs.
-  if (cc->Outputs().HasTag("MASK")) {
-    cc->Outputs().Tag("MASK").Set<ImageFrame>();
+  if (cc->Outputs().HasTag(kMaskTag)) {
+    cc->Outputs().Tag(kMaskTag).Set<ImageFrame>();
  }
 #if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
-  if (cc->Outputs().HasTag("MASK_GPU")) {
-    cc->Outputs().Tag("MASK_GPU").Set<mediapipe::GpuBuffer>();
+  if (cc->Outputs().HasTag(kMaskGpuTag)) {
+    cc->Outputs().Tag(kMaskGpuTag).Set<mediapipe::GpuBuffer>();
    use_gpu |= true;
  }
 #endif  //  !MEDIAPIPE_DISABLE_GPU
@@ -197,7 +207,7 @@ REGISTER_CALCULATOR(TfLiteTensorsToSegmentationCalculator);
    CalculatorContext* cc) {
  cc->SetOffset(TimestampDiff(0));

-  if (cc->Inputs().HasTag("TENSORS_GPU")) {
+  if (cc->Inputs().HasTag(kTensorsGpuTag)) {
    use_gpu_ = true;
 #if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
    MP_RETURN_IF_ERROR(gpu_helper_.Open(cc));
@@ -255,23 +265,22 @@ REGISTER_CALCULATOR(TfLiteTensorsToSegmentationCalculator);

 ::mediapipe::Status TfLiteTensorsToSegmentationCalculator::ProcessCpu(
    CalculatorContext* cc) {
-  if (cc->Inputs().Tag("TENSORS").IsEmpty()) {
+  if (cc->Inputs().Tag(kTensorsTag).IsEmpty()) {
    return ::mediapipe::OkStatus();
  }

  // Get input streams.
  const auto& input_tensors =
-      cc->Inputs().Tag("TENSORS").Get<std::vector<TfLiteTensor>>();
-  const bool has_prev_mask = cc->Inputs().HasTag("PREV_MASK") &&
-                             !cc->Inputs().Tag("PREV_MASK").IsEmpty();
+      cc->Inputs().Tag(kTensorsTag).Get<std::vector<TfLiteTensor>>();
+  const bool has_prev_mask = cc->Inputs().HasTag(kPrevMaskTag) &&
+                             !cc->Inputs().Tag(kPrevMaskTag).IsEmpty();
  const ImageFrame placeholder;
-  const auto& input_mask = has_prev_mask
-                               ? cc->Inputs().Tag("PREV_MASK").Get<ImageFrame>()
-                               : placeholder;
+  const auto& input_mask =
+      has_prev_mask ? cc->Inputs().Tag(kPrevMaskTag).Get<ImageFrame>()
+                    : placeholder;
  int output_width = tensor_width_, output_height = tensor_height_;
-  if (cc->Inputs().HasTag("REFERENCE_IMAGE")) {
-    const auto& input_image =
-        cc->Inputs().Tag("REFERENCE_IMAGE").Get<ImageFrame>();
+  if (cc->Inputs().HasTag(kSizeImageTag)) {
+    const auto& input_image = cc->Inputs().Tag(kSizeImageTag).Get<ImageFrame>();
    output_width = input_image.Width();
    output_height = input_image.Height();
  }
@@ -353,7 +362,7 @@ REGISTER_CALCULATOR(TfLiteTensorsToSegmentationCalculator);
      ImageFormat::SRGBA, output_width, output_height);
  cv::Mat output_mat = formats::MatView(output_mask.get());
  large_mask_mat.copyTo(output_mat);
-  cc->Outputs().Tag("MASK").Add(output_mask.release(), cc->InputTimestamp());
+  cc->Outputs().Tag(kMaskTag).Add(output_mask.release(), cc->InputTimestamp());

  return ::mediapipe::OkStatus();
 }
@@ -364,23 +373,23 @@ REGISTER_CALCULATOR(TfLiteTensorsToSegmentationCalculator);
 // 3. upsample small mask into output mask to be same size as input image
 ::mediapipe::Status TfLiteTensorsToSegmentationCalculator::ProcessGpu(
    CalculatorContext* cc) {
-  if (cc->Inputs().Tag("TENSORS_GPU").IsEmpty()) {
+  if (cc->Inputs().Tag(kTensorsGpuTag).IsEmpty()) {
    return ::mediapipe::OkStatus();
  }
 #if !defined(MEDIAPIPE_DISABLE_GL_COMPUTE)
  // Get input streams.
  const auto& input_tensors =
-      cc->Inputs().Tag("TENSORS_GPU").Get<std::vector<GlBuffer>>();
-  const bool has_prev_mask = cc->Inputs().HasTag("PREV_MASK_GPU") &&
-                             !cc->Inputs().Tag("PREV_MASK_GPU").IsEmpty();
+      cc->Inputs().Tag(kTensorsGpuTag).Get<std::vector<GlBuffer>>();
+  const bool has_prev_mask = cc->Inputs().HasTag(kPrevMaskGpuTag) &&
+                             !cc->Inputs().Tag(kPrevMaskGpuTag).IsEmpty();
  const auto& input_mask =
      has_prev_mask
-          ? cc->Inputs().Tag("PREV_MASK_GPU").Get<mediapipe::GpuBuffer>()
+          ? cc->Inputs().Tag(kPrevMaskGpuTag).Get<mediapipe::GpuBuffer>()
          : mediapipe::GpuBuffer();
  int output_width = tensor_width_, output_height = tensor_height_;
-  if (cc->Inputs().HasTag("REFERENCE_IMAGE_GPU")) {
+  if (cc->Inputs().HasTag(kSizeImageGpuTag)) {
    const auto& input_image =
-        cc->Inputs().Tag("REFERENCE_IMAGE_GPU").Get<mediapipe::GpuBuffer>();
+        cc->Inputs().Tag(kSizeImageGpuTag).Get<mediapipe::GpuBuffer>();
    output_width = input_image.width();
    output_height = input_image.height();
  }
@@ -441,7 +450,7 @@ REGISTER_CALCULATOR(TfLiteTensorsToSegmentationCalculator);
  // Send out image as GPU packet.
  auto output_image = output_texture.GetFrame<mediapipe::GpuBuffer>();
  cc->Outputs()
-      .Tag("MASK_GPU")
+      .Tag(kMaskGpuTag)
      .Add(output_image.release(), cc->InputTimestamp());

  // Cleanup
@@ -15,11 +15,11 @@
 #ifndef MEDIAPIPE_CALCULATORS_TFLITE_UTIL_H_
 #define MEDIAPIPE_CALCULATORS_TFLITE_UTIL_H_

-#define RET_CHECK_CALL(call)                                     \
-  do {                                                           \
-    const auto status = (call);                                  \
-    if (ABSL_PREDICT_FALSE(!status.ok()))                        \
-      return ::mediapipe::InternalError(status.error_message()); \
+#define RET_CHECK_CALL(call)                               \
+  do {                                                     \
+    const auto status = (call);                            \
+    if (ABSL_PREDICT_FALSE(!status.ok()))                  \
+      return ::mediapipe::InternalError(status.message()); \
  } while (0);

 #endif  // MEDIAPIPE_CALCULATORS_TFLITE_UTIL_H_
@@ -321,7 +321,7 @@ cc_library(
        "//mediapipe:android": [
            "//mediapipe/util/android/file/base",
        ],
-        "//mediapipe:apple": [
+        "//mediapipe:ios": [
            "//mediapipe/util/android/file/base",
        ],
        "//mediapipe:macos": [
@@ -349,7 +349,7 @@ cc_library(
        "//mediapipe:android": [
            "//mediapipe/util/android/file/base",
        ],
-        "//mediapipe:apple": [
+        "//mediapipe:ios": [
            "//mediapipe/util/android/file/base",
        ],
        "//mediapipe:macos": [
@@ -926,7 +926,7 @@ cc_library(
        "//mediapipe:android": [
            "//mediapipe/util/android/file/base",
        ],
-        "//mediapipe:apple": [
+        "//mediapipe:ios": [
            "//mediapipe/util/android/file/base",
        ],
        "//mediapipe:macos": [
@@ -971,9 +971,9 @@ cc_library(
    visibility = ["//visibility:public"],
    deps = [
        "//mediapipe/framework:calculator_framework",
-        "//mediapipe/framework/port:file_helpers",
        "//mediapipe/framework/port:ret_check",
        "//mediapipe/framework/port:status",
+        "//mediapipe/util:resource_util",
    ],
    alwayslink = 1,
 )
@@ -998,6 +998,7 @@ cc_library(
    visibility = ["//visibility:public"],
    deps = [
        "//mediapipe/framework:calculator_framework",
+        "//mediapipe/framework/formats:classification_cc_proto",
        "//mediapipe/framework/formats:landmark_cc_proto",
        "//mediapipe/framework/formats:rect_cc_proto",
        "//mediapipe/framework/port:ret_check",
@@ -1015,6 +1016,7 @@ cc_library(
    deps = [
        ":collection_has_min_size_calculator_cc_proto",
        "//mediapipe/framework:calculator_framework",
+        "//mediapipe/framework/formats:landmark_cc_proto",
        "//mediapipe/framework/formats:rect_cc_proto",
        "//mediapipe/framework/port:ret_check",
        "//mediapipe/framework/port:status",
@@ -1022,6 +1024,18 @@ cc_library(
    alwayslink = 1,
 )

+cc_test(
+    name = "collection_has_min_size_calculator_test",
+    srcs = ["collection_has_min_size_calculator_test.cc"],
+    deps = [
+        ":collection_has_min_size_calculator",
+        "//mediapipe/framework:calculator_framework",
+        "//mediapipe/framework:calculator_runner",
+        "//mediapipe/framework/port:gtest_main",
+        "//mediapipe/framework/port:parse_text_proto",
+    ],
+)
+
 cc_library(
    name = "association_calculator",
    hdrs = ["association_calculator.h"],
@@ -55,7 +55,7 @@ size_t RoundUp(size_t n, size_t m) { return ((n + m - 1) / m) * m; }  // NOLINT
 // When using GPU, this color will become transparent when the calculator
 // merges the annotation overlay with the image frame. As a result, drawing in
 // this color is not supported and it should be set to something unlikely used.
-constexpr int kAnnotationBackgroundColor[] = {100, 101, 102};
+constexpr uchar kAnnotationBackgroundColor = 2;  // Grayscale value.
 }  // namespace

 // A calculator for rendering data on images.
@@ -491,11 +491,9 @@ REGISTER_CALCULATOR(AnnotationOverlayCalculator);
    if (format != mediapipe::ImageFormat::SRGBA &&
        format != mediapipe::ImageFormat::SRGB)
      RET_CHECK_FAIL() << "Unsupported GPU input format: " << format;
-
-    image_mat = absl::make_unique<cv::Mat>(
-        height_, width_, CV_8UC3,
-        cv::Scalar(kAnnotationBackgroundColor[0], kAnnotationBackgroundColor[1],
-                   kAnnotationBackgroundColor[2]));
+    image_mat = absl::make_unique<cv::Mat>(height_, width_, CV_8UC3);
+    memset(image_mat->data, kAnnotationBackgroundColor,
+           height_ * width_ * image_mat->elemSize());
  } else {
    image_mat = absl::make_unique<cv::Mat>(
        options_.canvas_height_px(), options_.canvas_width_px(), CV_8UC3,
@@ -617,9 +615,9 @@ REGISTER_CALCULATOR(AnnotationOverlayCalculator);
  glUniform1i(glGetUniformLocation(program_, "input_frame"), 1);
  glUniform1i(glGetUniformLocation(program_, "overlay"), 2);
  glUniform3f(glGetUniformLocation(program_, "transparent_color"),
-              kAnnotationBackgroundColor[0] / 255.0,
-              kAnnotationBackgroundColor[1] / 255.0,
-              kAnnotationBackgroundColor[2] / 255.0);
+              kAnnotationBackgroundColor / 255.0,
+              kAnnotationBackgroundColor / 255.0,
+              kAnnotationBackgroundColor / 255.0);

  // Init texture for opencv rendered frame.
  const auto& input_frame =
@@ -15,6 +15,9 @@

 #include "mediapipe/calculators/util/collection_has_min_size_calculator.h"

+#include <vector>
+
+#include "mediapipe/framework/formats/landmark.pb.h"
 #include "mediapipe/framework/formats/rect.pb.h"

 namespace mediapipe {
@@ -23,4 +26,9 @@ typedef CollectionHasMinSizeCalculator<std::vector<::mediapipe::NormalizedRect>>
    NormalizedRectVectorHasMinSizeCalculator;
 REGISTER_CALCULATOR(NormalizedRectVectorHasMinSizeCalculator);

+typedef CollectionHasMinSizeCalculator<
+    std::vector<::mediapipe::NormalizedLandmarkList>>
+    NormalizedLandmarkListVectorHasMinSizeCalculator;
+REGISTER_CALCULATOR(NormalizedLandmarkListVectorHasMinSizeCalculator);
+
 }  // namespace mediapipe
@@ -0,0 +1,156 @@
+// Copyright 2020 The MediaPipe Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "mediapipe/calculators/util/collection_has_min_size_calculator.h"
+
+#include <memory>
+#include <string>
+#include <vector>
+
+#include "mediapipe/framework/calculator_framework.h"
+#include "mediapipe/framework/calculator_runner.h"
+#include "mediapipe/framework/port/gmock.h"
+#include "mediapipe/framework/port/gtest.h"
+#include "mediapipe/framework/port/parse_text_proto.h"
+#include "mediapipe/framework/port/status_matchers.h"  // NOLINT
+
+namespace mediapipe {
+
+typedef CollectionHasMinSizeCalculator<std::vector<int>>
+    TestIntCollectionHasMinSizeCalculator;
+REGISTER_CALCULATOR(TestIntCollectionHasMinSizeCalculator);
+
+void AddInputVector(const std::vector<int>& input, int64 timestamp,
+                    CalculatorRunner* runner) {
+  runner->MutableInputs()
+      ->Tag("ITERABLE")
+      .packets.push_back(
+          MakePacket<std::vector<int>>(input).At(Timestamp(timestamp)));
+}
+
+TEST(TestIntCollectionHasMinSizeCalculator, DoesHaveMinSize) {
+  CalculatorGraphConfig::Node node_config =
+      ParseTextProtoOrDie<CalculatorGraphConfig::Node>(R"(
+        calculator: "TestIntCollectionHasMinSizeCalculator"
+        input_stream: "ITERABLE:input_vector"
+        output_stream: "output_vector"
+        options {
+          [mediapipe.CollectionHasMinSizeCalculatorOptions.ext] { min_size: 2 }
+        }
+      )");
+  CalculatorRunner runner(node_config);
+  const std::vector<Packet>& outputs = runner.Outputs().Index(0).packets;
+
+  AddInputVector({1, 2}, /*timestamp=*/1, &runner);
+  MP_ASSERT_OK(runner.Run());
+
+  EXPECT_EQ(1, outputs.size());
+  EXPECT_EQ(Timestamp(1), outputs[0].Timestamp());
+  EXPECT_TRUE(outputs[0].Get<bool>());
+
+  AddInputVector({1, 2, 3}, /*timestamp=*/2, &runner);
+  MP_ASSERT_OK(runner.Run());
+
+  EXPECT_EQ(2, outputs.size());
+  EXPECT_EQ(Timestamp(2), outputs[1].Timestamp());
+  EXPECT_TRUE(outputs[1].Get<bool>());
+}
+
+TEST(TestIntCollectionHasMinSizeCalculator,
+     DoesHaveMinSize_MinSizeAsSidePacket) {
+  CalculatorGraphConfig::Node node_config =
+      ParseTextProtoOrDie<CalculatorGraphConfig::Node>(R"(
+        calculator: "TestIntCollectionHasMinSizeCalculator"
+        input_stream: "ITERABLE:input_vector"
+        input_side_packet: "min_size"
+        output_stream: "output_vector"
+      )");
+  CalculatorRunner runner(node_config);
+  const std::vector<Packet>& outputs = runner.Outputs().Index(0).packets;
+
+  runner.MutableSidePackets()->Index(0) = MakePacket<int>(2);
+
+  AddInputVector({1, 2}, /*timestamp=*/1, &runner);
+  MP_ASSERT_OK(runner.Run());
+
+  EXPECT_EQ(1, outputs.size());
+  EXPECT_EQ(Timestamp(1), outputs[0].Timestamp());
+  EXPECT_TRUE(outputs[0].Get<bool>());
+
+  AddInputVector({1, 2, 3}, /*timestamp=*/2, &runner);
+  MP_ASSERT_OK(runner.Run());
+
+  EXPECT_EQ(2, outputs.size());
+  EXPECT_EQ(Timestamp(2), outputs[1].Timestamp());
+  EXPECT_TRUE(outputs[1].Get<bool>());
+}
+
+TEST(TestIntCollectionHasMinSizeCalculator, DoesNotHaveMinSize) {
+  CalculatorGraphConfig::Node node_config =
+      ParseTextProtoOrDie<CalculatorGraphConfig::Node>(R"(
+        calculator: "TestIntCollectionHasMinSizeCalculator"
+        input_stream: "ITERABLE:input_vector"
+        output_stream: "output_vector"
+        options {
+          [mediapipe.CollectionHasMinSizeCalculatorOptions.ext] { min_size: 3 }
+        }
+      )");
+  CalculatorRunner runner(node_config);
+  const std::vector<Packet>& outputs = runner.Outputs().Index(0).packets;
+
+  AddInputVector({1}, /*timestamp=*/1, &runner);
+  MP_ASSERT_OK(runner.Run());
+
+  EXPECT_EQ(1, outputs.size());
+  EXPECT_EQ(Timestamp(1), outputs[0].Timestamp());
+  EXPECT_FALSE(outputs[0].Get<bool>());
+
+  AddInputVector({1, 2}, /*timestamp=*/2, &runner);
+  MP_ASSERT_OK(runner.Run());
+
+  EXPECT_EQ(2, outputs.size());
+  EXPECT_EQ(Timestamp(2), outputs[1].Timestamp());
+  EXPECT_FALSE(outputs[1].Get<bool>());
+}
+
+TEST(TestIntCollectionHasMinSizeCalculator,
+     DoesNotHaveMinSize_MinSizeAsSidePacket) {
+  CalculatorGraphConfig::Node node_config =
+      ParseTextProtoOrDie<CalculatorGraphConfig::Node>(R"(
+        calculator: "TestIntCollectionHasMinSizeCalculator"
+        input_stream: "ITERABLE:input_vector"
+        input_side_packet: "min_size"
+        output_stream: "output_vector"
+      )");
+  CalculatorRunner runner(node_config);
+  const std::vector<Packet>& outputs = runner.Outputs().Index(0).packets;
+
+  runner.MutableSidePackets()->Index(0) = MakePacket<int>(3);
+
+  AddInputVector({1}, /*timestamp=*/1, &runner);
+  MP_ASSERT_OK(runner.Run());
+
+  EXPECT_EQ(1, outputs.size());
+  EXPECT_EQ(Timestamp(1), outputs[0].Timestamp());
+  EXPECT_FALSE(outputs[0].Get<bool>());
+
+  AddInputVector({1, 2}, /*timestamp=*/2, &runner);
+  MP_ASSERT_OK(runner.Run());
+
+  EXPECT_EQ(2, outputs.size());
+  EXPECT_EQ(Timestamp(2), outputs[1].Timestamp());
+  EXPECT_FALSE(outputs[1].Get<bool>());
+}
+
+}  // namespace mediapipe
@@ -71,16 +71,23 @@ REGISTER_CALCULATOR(DetectionLabelIdToTextCalculator);
  const auto& options =
      cc->Options<::mediapipe::DetectionLabelIdToTextCalculatorOptions>();

-  std::string string_path;
-  ASSIGN_OR_RETURN(string_path, PathToResourceAsFile(options.label_map_path()));
-  std::string label_map_string;
-  MP_RETURN_IF_ERROR(file::GetContents(string_path, &label_map_string));
+  if (options.has_label_map_path()) {
+    std::string string_path;
+    ASSIGN_OR_RETURN(string_path,
+                     PathToResourceAsFile(options.label_map_path()));
+    std::string label_map_string;
+    MP_RETURN_IF_ERROR(file::GetContents(string_path, &label_map_string));

-  std::istringstream stream(label_map_string);
-  std::string line;
-  int i = 0;
-  while (std::getline(stream, line)) {
-    label_map_[i++] = line;
+    std::istringstream stream(label_map_string);
+    std::string line;
+    int i = 0;
+    while (std::getline(stream, line)) {
+      label_map_[i++] = line;
+    }
+  } else {
+    for (int i = 0; i < options.label_size(); ++i) {
+      label_map_[i] = options.label(i);
+    }
  }
  return ::mediapipe::OkStatus();
 }
@@ -25,4 +25,10 @@ message DetectionLabelIdToTextCalculatorOptions {

  // Path to a label map file for getting the actual name of detected classes.
  optional string label_map_path = 1;
+
+  // Alternative way to specify label map
+  // label: "label for id 0"
+  // label: "label for id 1"
+  // ...
+  repeated string label = 2;
 }
@@ -17,6 +17,7 @@

 #include <vector>

+#include "mediapipe/framework/formats/classification.pb.h"
 #include "mediapipe/framework/formats/landmark.pb.h"
 #include "mediapipe/framework/formats/rect.pb.h"

@@ -31,4 +32,8 @@ typedef FilterCollectionCalculator<
    FilterLandmarkListCollectionCalculator;
 REGISTER_CALCULATOR(FilterLandmarkListCollectionCalculator);

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

@@ -29,6 +29,7 @@ namespace {

 constexpr char kLandmarksTag[] = "LANDMARKS";
 constexpr char kNormLandmarksTag[] = "NORM_LANDMARKS";
+constexpr char kRenderScaleTag[] = "RENDER_SCALE";
 constexpr char kRenderDataTag[] = "RENDER_DATA";
 constexpr char kLandmarkLabel[] = "KEYPOINT";
 constexpr int kMaxLandmarkThickness = 18;
@@ -71,6 +72,83 @@ void SetColorSizeValueFromZ(float z, float z_min, float z_max,
  render_annotation->set_thickness(thickness);
 }

+template <class LandmarkType>
+void AddConnectionToRenderData(const LandmarkType& start,
+                               const LandmarkType& end, int gray_val1,
+                               int gray_val2, float thickness, bool normalized,
+                               RenderData* render_data) {
+  auto* connection_annotation = render_data->add_render_annotations();
+  RenderAnnotation::GradientLine* line =
+      connection_annotation->mutable_gradient_line();
+  line->set_x_start(start.x());
+  line->set_y_start(start.y());
+  line->set_x_end(end.x());
+  line->set_y_end(end.y());
+  line->set_normalized(normalized);
+  line->mutable_color1()->set_r(gray_val1);
+  line->mutable_color1()->set_g(gray_val1);
+  line->mutable_color1()->set_b(gray_val1);
+  line->mutable_color2()->set_r(gray_val2);
+  line->mutable_color2()->set_g(gray_val2);
+  line->mutable_color2()->set_b(gray_val2);
+  connection_annotation->set_thickness(thickness);
+}
+
+template <class LandmarkListType, class LandmarkType>
+void AddConnectionsWithDepth(const LandmarkListType& landmarks,
+                             const std::vector<int>& landmark_connections,
+                             float thickness, bool normalized, float min_z,
+                             float max_z, RenderData* render_data) {
+  for (int i = 0; i < landmark_connections.size(); i += 2) {
+    const auto& ld0 = landmarks.landmark(landmark_connections[i]);
+    const auto& ld1 = landmarks.landmark(landmark_connections[i + 1]);
+    const int gray_val1 =
+        255 - static_cast<int>(Remap(ld0.z(), min_z, max_z, 255));
+    const int gray_val2 =
+        255 - static_cast<int>(Remap(ld1.z(), min_z, max_z, 255));
+    AddConnectionToRenderData<LandmarkType>(ld0, ld1, gray_val1, gray_val2,
+                                            thickness, normalized, render_data);
+  }
+}
+
+template <class LandmarkType>
+void AddConnectionToRenderData(const LandmarkType& start,
+                               const LandmarkType& end,
+                               const Color& connection_color, float thickness,
+                               bool normalized, RenderData* render_data) {
+  auto* connection_annotation = render_data->add_render_annotations();
+  RenderAnnotation::Line* line = connection_annotation->mutable_line();
+  line->set_x_start(start.x());
+  line->set_y_start(start.y());
+  line->set_x_end(end.x());
+  line->set_y_end(end.y());
+  line->set_normalized(normalized);
+  SetColor(connection_annotation, connection_color);
+  connection_annotation->set_thickness(thickness);
+}
+
+template <class LandmarkListType, class LandmarkType>
+void AddConnections(const LandmarkListType& landmarks,
+                    const std::vector<int>& landmark_connections,
+                    const Color& connection_color, float thickness,
+                    bool normalized, RenderData* render_data) {
+  for (int i = 0; i < landmark_connections.size(); i += 2) {
+    const auto& ld0 = landmarks.landmark(landmark_connections[i]);
+    const auto& ld1 = landmarks.landmark(landmark_connections[i + 1]);
+    AddConnectionToRenderData<LandmarkType>(ld0, ld1, connection_color,
+                                            thickness, normalized, render_data);
+  }
+}
+
+RenderAnnotation* AddPointRenderData(const Color& landmark_color,
+                                     float thickness, RenderData* render_data) {
+  auto* landmark_data_annotation = render_data->add_render_annotations();
+  landmark_data_annotation->set_scene_tag(kLandmarkLabel);
+  SetColor(landmark_data_annotation, landmark_color);
+  landmark_data_annotation->set_thickness(thickness);
+  return landmark_data_annotation;
+}
+
 }  // namespace

 // A calculator that converts Landmark proto to RenderData proto for
@@ -107,30 +185,8 @@ class LandmarksToRenderDataCalculator : public CalculatorBase {
  ::mediapipe::Status Process(CalculatorContext* cc) override;

 private:
-  static void AddConnectionToRenderData(
-      float start_x, float start_y, float end_x, float end_y,
-      const LandmarksToRenderDataCalculatorOptions& options, bool normalized,
-      RenderData* render_data);
-  static void SetRenderAnnotationColorThickness(
-      const LandmarksToRenderDataCalculatorOptions& options,
-      RenderAnnotation* render_annotation);
-  static RenderAnnotation* AddPointRenderData(
-      const LandmarksToRenderDataCalculatorOptions& options,
-      RenderData* render_data);
-  static void AddConnectionToRenderData(
-      float start_x, float start_y, float end_x, float end_y,
-      const LandmarksToRenderDataCalculatorOptions& options, bool normalized,
-      int gray_val1, int gray_val2, RenderData* render_data);
-
-  template <class LandmarkListType>
-  void AddConnections(const LandmarkListType& landmarks, bool normalized,
-                      RenderData* render_data);
-  template <class LandmarkListType>
-  void AddConnectionsWithDepth(const LandmarkListType& landmarks,
-                               bool normalized, float min_z, float max_z,
-                               RenderData* render_data);
-
  LandmarksToRenderDataCalculatorOptions options_;
+  std::vector<int> landmark_connections_;
 };
 REGISTER_CALCULATOR(LandmarksToRenderDataCalculator);

@@ -150,6 +206,9 @@ REGISTER_CALCULATOR(LandmarksToRenderDataCalculator);
  if (cc->Inputs().HasTag(kNormLandmarksTag)) {
    cc->Inputs().Tag(kNormLandmarksTag).Set<NormalizedLandmarkList>();
  }
+  if (cc->Inputs().HasTag(kRenderScaleTag)) {
+    cc->Inputs().Tag(kRenderScaleTag).Set<float>();
+  }
  cc->Outputs().Tag(kRenderDataTag).Set<RenderData>();
  return ::mediapipe::OkStatus();
 }
@@ -159,6 +218,14 @@ REGISTER_CALCULATOR(LandmarksToRenderDataCalculator);
  cc->SetOffset(TimestampDiff(0));
  options_ = cc->Options<LandmarksToRenderDataCalculatorOptions>();

+  // Parse landmarks connections to a vector.
+  RET_CHECK_EQ(options_.landmark_connections_size() % 2, 0)
+      << "Number of entries in landmark connections must be a multiple of 2";
+
+  for (int i = 0; i < options_.landmark_connections_size(); ++i) {
+    landmark_connections_.push_back(options_.landmark_connections(i));
+  }
+
  return ::mediapipe::OkStatus();
 }

@@ -169,20 +236,36 @@ REGISTER_CALCULATOR(LandmarksToRenderDataCalculator);
  float z_min = 0.f;
  float z_max = 0.f;

+  // Apply scale to `thickness` of rendered landmarks and connections to make
+  // them bigger when object (e.g. pose, hand or face) is closer/bigger and
+  // snaller when object is further/smaller.
+  float thickness = options_.thickness();
+  if (cc->Inputs().HasTag(kRenderScaleTag)) {
+    const float render_scale = cc->Inputs().Tag(kRenderScaleTag).Get<float>();
+    thickness *= render_scale;
+  }
+
  if (cc->Inputs().HasTag(kLandmarksTag)) {
    const LandmarkList& landmarks =
        cc->Inputs().Tag(kLandmarksTag).Get<LandmarkList>();
-    RET_CHECK_EQ(options_.landmark_connections_size() % 2, 0)
-        << "Number of entries in landmark connections must be a multiple of 2";
    if (visualize_depth) {
      GetMinMaxZ<LandmarkList, Landmark>(landmarks, &z_min, &z_max);
    }
    // Only change rendering if there are actually z values other than 0.
    visualize_depth &= ((z_max - z_min) > 1e-3);
+    if (visualize_depth) {
+      AddConnectionsWithDepth<LandmarkList, Landmark>(
+          landmarks, landmark_connections_, thickness, /*normalized=*/false,
+          z_min, z_max, render_data.get());
+    } else {
+      AddConnections<LandmarkList, Landmark>(
+          landmarks, landmark_connections_, options_.connection_color(),
+          thickness, /*normalized=*/false, render_data.get());
+    }
    for (int i = 0; i < landmarks.landmark_size(); ++i) {
      const Landmark& landmark = landmarks.landmark(i);
-      auto* landmark_data_render =
-          AddPointRenderData(options_, render_data.get());
+      auto* landmark_data_render = AddPointRenderData(
+          options_.landmark_color(), thickness, render_data.get());
      if (visualize_depth) {
        SetColorSizeValueFromZ(landmark.z(), z_min, z_max,
                               landmark_data_render);
@@ -192,30 +275,30 @@ REGISTER_CALCULATOR(LandmarksToRenderDataCalculator);
      landmark_data->set_x(landmark.x());
      landmark_data->set_y(landmark.y());
    }
-    if (visualize_depth) {
-      AddConnectionsWithDepth<LandmarkList>(landmarks, /*normalized=*/false,
-                                            z_min, z_max, render_data.get());
-    } else {
-      AddConnections<LandmarkList>(landmarks, /*normalized=*/false,
-                                   render_data.get());
-    }
  }

  if (cc->Inputs().HasTag(kNormLandmarksTag)) {
    const NormalizedLandmarkList& landmarks =
        cc->Inputs().Tag(kNormLandmarksTag).Get<NormalizedLandmarkList>();
-    RET_CHECK_EQ(options_.landmark_connections_size() % 2, 0)
-        << "Number of entries in landmark connections must be a multiple of 2";
    if (visualize_depth) {
      GetMinMaxZ<NormalizedLandmarkList, NormalizedLandmark>(landmarks, &z_min,
                                                             &z_max);
    }
    // Only change rendering if there are actually z values other than 0.
    visualize_depth &= ((z_max - z_min) > 1e-3);
+    if (visualize_depth) {
+      AddConnectionsWithDepth<NormalizedLandmarkList, NormalizedLandmark>(
+          landmarks, landmark_connections_, thickness, /*normalized=*/true,
+          z_min, z_max, render_data.get());
+    } else {
+      AddConnections<NormalizedLandmarkList, NormalizedLandmark>(
+          landmarks, landmark_connections_, options_.connection_color(),
+          thickness, /*normalized=*/true, render_data.get());
+    }
    for (int i = 0; i < landmarks.landmark_size(); ++i) {
      const NormalizedLandmark& landmark = landmarks.landmark(i);
-      auto* landmark_data_render =
-          AddPointRenderData(options_, render_data.get());
+      auto* landmark_data_render = AddPointRenderData(
+          options_.landmark_color(), thickness, render_data.get());
      if (visualize_depth) {
        SetColorSizeValueFromZ(landmark.z(), z_min, z_max,
                               landmark_data_render);
@@ -225,13 +308,6 @@ REGISTER_CALCULATOR(LandmarksToRenderDataCalculator);
      landmark_data->set_x(landmark.x());
      landmark_data->set_y(landmark.y());
    }
-    if (visualize_depth) {
-      AddConnectionsWithDepth<NormalizedLandmarkList>(
-          landmarks, /*normalized=*/true, z_min, z_max, render_data.get());
-    } else {
-      AddConnections<NormalizedLandmarkList>(landmarks, /*normalized=*/true,
-                                             render_data.get());
-    }
  }

  cc->Outputs()
@@ -240,84 +316,4 @@ REGISTER_CALCULATOR(LandmarksToRenderDataCalculator);
  return ::mediapipe::OkStatus();
 }

-template <class LandmarkListType>
-void LandmarksToRenderDataCalculator::AddConnectionsWithDepth(
-    const LandmarkListType& landmarks, bool normalized, float min_z,
-    float max_z, RenderData* render_data) {
-  for (int i = 0; i < options_.landmark_connections_size(); i += 2) {
-    const auto& ld0 = landmarks.landmark(options_.landmark_connections(i));
-    const auto& ld1 = landmarks.landmark(options_.landmark_connections(i + 1));
-    const int gray_val1 =
-        255 - static_cast<int>(Remap(ld0.z(), min_z, max_z, 255));
-    const int gray_val2 =
-        255 - static_cast<int>(Remap(ld1.z(), min_z, max_z, 255));
-    AddConnectionToRenderData(ld0.x(), ld0.y(), ld1.x(), ld1.y(), options_,
-                              normalized, gray_val1, gray_val2, render_data);
-  }
-}
-
-void LandmarksToRenderDataCalculator::AddConnectionToRenderData(
-    float start_x, float start_y, float end_x, float end_y,
-    const LandmarksToRenderDataCalculatorOptions& options, bool normalized,
-    int gray_val1, int gray_val2, RenderData* render_data) {
-  auto* connection_annotation = render_data->add_render_annotations();
-  RenderAnnotation::GradientLine* line =
-      connection_annotation->mutable_gradient_line();
-  line->set_x_start(start_x);
-  line->set_y_start(start_y);
-  line->set_x_end(end_x);
-  line->set_y_end(end_y);
-  line->set_normalized(normalized);
-  line->mutable_color1()->set_r(gray_val1);
-  line->mutable_color1()->set_g(gray_val1);
-  line->mutable_color1()->set_b(gray_val1);
-  line->mutable_color2()->set_r(gray_val2);
-  line->mutable_color2()->set_g(gray_val2);
-  line->mutable_color2()->set_b(gray_val2);
-  connection_annotation->set_thickness(options.thickness());
-}
-
-template <class LandmarkListType>
-void LandmarksToRenderDataCalculator::AddConnections(
-    const LandmarkListType& landmarks, bool normalized,
-    RenderData* render_data) {
-  for (int i = 0; i < options_.landmark_connections_size(); i += 2) {
-    const auto& ld0 = landmarks.landmark(options_.landmark_connections(i));
-    const auto& ld1 = landmarks.landmark(options_.landmark_connections(i + 1));
-    AddConnectionToRenderData(ld0.x(), ld0.y(), ld1.x(), ld1.y(), options_,
-                              normalized, render_data);
-  }
-}
-
-void LandmarksToRenderDataCalculator::AddConnectionToRenderData(
-    float start_x, float start_y, float end_x, float end_y,
-    const LandmarksToRenderDataCalculatorOptions& options, bool normalized,
-    RenderData* render_data) {
-  auto* connection_annotation = render_data->add_render_annotations();
-  RenderAnnotation::Line* line = connection_annotation->mutable_line();
-  line->set_x_start(start_x);
-  line->set_y_start(start_y);
-  line->set_x_end(end_x);
-  line->set_y_end(end_y);
-  line->set_normalized(normalized);
-  SetColor(connection_annotation, options.connection_color());
-  connection_annotation->set_thickness(options.thickness());
-}
-
-RenderAnnotation* LandmarksToRenderDataCalculator::AddPointRenderData(
-    const LandmarksToRenderDataCalculatorOptions& options,
-    RenderData* render_data) {
-  auto* landmark_data_annotation = render_data->add_render_annotations();
-  landmark_data_annotation->set_scene_tag(kLandmarkLabel);
-  SetRenderAnnotationColorThickness(options, landmark_data_annotation);
-  return landmark_data_annotation;
-}
-
-void LandmarksToRenderDataCalculator::SetRenderAnnotationColorThickness(
-    const LandmarksToRenderDataCalculatorOptions& options,
-    RenderAnnotation* render_annotation) {
-  SetColor(render_annotation, options.landmark_color());
-  render_annotation->set_thickness(options.thickness());
-}
-
 }  // namespace mediapipe
@@ -16,34 +16,80 @@
 #include <string>

 #include "mediapipe/framework/calculator_framework.h"
-#include "mediapipe/framework/port/file_helpers.h"
 #include "mediapipe/framework/port/status.h"
+#include "mediapipe/util/resource_util.h"

 namespace mediapipe {
+
+namespace {
+
+constexpr char kFilePathTag[] = "FILE_PATH";
+constexpr char kContentsTag[] = "CONTENTS";
+
+}  // namespace
+
 // The calculator takes the path to the local file as an input side packet and
 // outputs the contents of that file.
 //
+// NOTE: file loading can be batched by providing multiple input/output side
+// packets.
+//
 // Example config:
 // node {
 //   calculator: "LocalFileContentsCalculator"
 //   input_side_packet: "FILE_PATH:file_path"
 //   output_side_packet: "CONTENTS:contents"
 // }
+//
+// node {
+//   calculator: "LocalFileContentsCalculator"
+//   input_side_packet: "FILE_PATH:0:file_path1"
+//   input_side_packet: "FILE_PATH:1:file_path2"
+//   ...
+//   output_side_packet: "CONTENTS:0:contents1"
+//   output_side_packet: "CONTENTS:1:contents2"
+//   ...
+// }
 class LocalFileContentsCalculator : public CalculatorBase {
 public:
  static ::mediapipe::Status GetContract(CalculatorContract* cc) {
-    cc->InputSidePackets().Tag("FILE_PATH").Set<std::string>();
-    cc->OutputSidePackets().Tag("CONTENTS").Set<std::string>();
+    RET_CHECK(cc->InputSidePackets().HasTag(kFilePathTag))
+        << "Missing PATH input side packet(s)";
+    RET_CHECK(cc->OutputSidePackets().HasTag(kContentsTag))
+        << "Missing CONTENTS output side packet(s)";
+
+    RET_CHECK_EQ(cc->InputSidePackets().NumEntries(kFilePathTag),
+                 cc->OutputSidePackets().NumEntries(kContentsTag))
+        << "Same number of input streams and output streams is required.";
+
+    for (CollectionItemId id = cc->InputSidePackets().BeginId(kFilePathTag);
+         id != cc->InputSidePackets().EndId(kFilePathTag); ++id) {
+      cc->InputSidePackets().Get(id).Set<std::string>();
+    }
+
+    for (CollectionItemId id = cc->OutputSidePackets().BeginId(kContentsTag);
+         id != cc->OutputSidePackets().EndId(kContentsTag); ++id) {
+      cc->OutputSidePackets().Get(id).Set<std::string>();
+    }
+
    return ::mediapipe::OkStatus();
  }

  ::mediapipe::Status Open(CalculatorContext* cc) override {
-    std::string contents;
-    MP_RETURN_IF_ERROR(mediapipe::file::GetContents(
-        cc->InputSidePackets().Tag("FILE_PATH").Get<std::string>(), &contents));
-    cc->OutputSidePackets()
-        .Tag("CONTENTS")
-        .Set(MakePacket<std::string>(std::move(contents)));
+    CollectionItemId input_id = cc->InputSidePackets().BeginId(kFilePathTag);
+    CollectionItemId output_id = cc->OutputSidePackets().BeginId(kContentsTag);
+    // Number of inputs and outpus is the same according to the contract.
+    for (; input_id != cc->InputSidePackets().EndId(kFilePathTag);
+         ++input_id, ++output_id) {
+      std::string file_path =
+          cc->InputSidePackets().Get(input_id).Get<std::string>();
+      ASSIGN_OR_RETURN(file_path, PathToResourceAsFile(file_path));
+
+      std::string contents;
+      MP_RETURN_IF_ERROR(GetResourceContents(file_path, &contents));
+      cc->OutputSidePackets().Get(output_id).Set(
+          MakePacket<std::string>(std::move(contents)));
+    }
    return ::mediapipe::OkStatus();
  }

@@ -276,6 +276,7 @@ TEST_F(PacketLatencyCalculatorTest, DoesNotOutputUntilReferencePacketReceived) {
      "delayed_packet_0", Adopt(new double()).At(Timestamp(2))));

  // Send a reference packet with timestamp 10 usec.
+  simulation_clock_->Sleep(absl::Microseconds(1));
  MP_ASSERT_OK(graph_.AddPacketToInputStream(
      "camera_frames", Adopt(new double()).At(Timestamp(10))));
  simulation_clock_->Sleep(absl::Microseconds(1));
@@ -12,6 +12,8 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.

+#include <algorithm>
+
 #include "absl/memory/memory.h"
 #include "absl/strings/str_cat.h"
 #include "absl/strings/str_join.h"
@@ -76,14 +78,16 @@ void AddTimedBoxProtoToRenderData(
    RenderAnnotation::Text* text = label_annotation->mutable_text();
    text->set_display_text(box_proto.label());
    text->set_normalized(true);
-    constexpr float text_left_start = 0.3f;
+    constexpr float text_left_start = 0.2f;
    text->set_left((1.0f - text_left_start) * box_proto.left() +
                   text_left_start * box_proto.right());
    constexpr float text_baseline = 0.6f;
    text->set_baseline(text_baseline * box_proto.bottom() +
                       (1.0f - text_baseline) * box_proto.top());
-    constexpr float text_height = 0.2f;
-    text->set_font_height((box_proto.bottom() - box_proto.top()) * text_height);
+    constexpr float text_height = 0.1f;
+    text->set_font_height(std::min(box_proto.bottom() - box_proto.top(),
+                                   box_proto.right() - box_proto.left()) *
+                          text_height);
  }
 }

@@ -65,6 +65,26 @@ proto_library(
    ],
 )

+proto_library(
+    name = "tracked_detection_manager_calculator_proto",
+    srcs = ["tracked_detection_manager_calculator.proto"],
+    visibility = ["//visibility:public"],
+    deps = [
+        "//mediapipe/framework:calculator_proto",
+        "//mediapipe/util/tracking:tracked_detection_manager_config_proto",
+    ],
+)
+
+proto_library(
+    name = "box_detector_calculator_proto",
+    srcs = ["box_detector_calculator.proto"],
+    visibility = ["//visibility:public"],
+    deps = [
+        "//mediapipe/framework:calculator_proto",
+        "//mediapipe/util/tracking:box_detector_proto",
+    ],
+)
+
 proto_library(
    name = "video_pre_stream_calculator_proto",
    srcs = ["video_pre_stream_calculator.proto"],
@@ -107,6 +127,28 @@ mediapipe_cc_proto_library(
    deps = [":box_tracker_calculator_proto"],
 )

+mediapipe_cc_proto_library(
+    name = "tracked_detection_manager_calculator_cc_proto",
+    srcs = ["tracked_detection_manager_calculator.proto"],
+    cc_deps = [
+        "//mediapipe/framework:calculator_cc_proto",
+        "//mediapipe/util/tracking:tracked_detection_manager_config_cc_proto",
+    ],
+    visibility = ["//visibility:public"],
+    deps = [":tracked_detection_manager_calculator_proto"],
+)
+
+mediapipe_cc_proto_library(
+    name = "box_detector_calculator_cc_proto",
+    srcs = ["box_detector_calculator.proto"],
+    cc_deps = [
+        "//mediapipe/framework:calculator_cc_proto",
+        "//mediapipe/util/tracking:box_detector_cc_proto",
+    ],
+    visibility = ["//visibility:public"],
+    deps = [":box_detector_calculator_proto"],
+)
+
 mediapipe_cc_proto_library(
    name = "video_pre_stream_calculator_cc_proto",
    srcs = ["video_pre_stream_calculator.proto"],
@@ -138,7 +180,7 @@ cc_library(
    srcs = ["flow_to_image_calculator.cc"],
    visibility = ["//visibility:public"],
    deps = [
-        "//mediapipe/calculators/video:flow_to_image_calculator_cc_proto",
+        ":flow_to_image_calculator_cc_proto",
        "//mediapipe/calculators/video/tool:flow_quantizer_model",
        "//mediapipe/framework:calculator_framework",
        "//mediapipe/framework/formats:image_format_cc_proto",
@@ -279,11 +321,54 @@ cc_library(
    alwayslink = 1,
 )

+cc_library(
+    name = "box_detector_calculator",
+    srcs = ["box_detector_calculator.cc"],
+    visibility = ["//visibility:public"],
+    deps = [
+        ":box_detector_calculator_cc_proto",
+        "@com_google_absl//absl/memory",
+        "@com_google_absl//absl/strings",
+        "//mediapipe/framework:calculator_framework",
+        "//mediapipe/framework/formats:image_frame",
+        "//mediapipe/framework/formats:image_frame_opencv",
+        "//mediapipe/framework/formats:video_stream_header",  # fixdeps: keep -- required for exobazel build.
+        "//mediapipe/framework/port:integral_types",
+        "//mediapipe/framework/port:logging",
+        "//mediapipe/framework/port:opencv_core",
+        "//mediapipe/framework/port:opencv_features2d",
+        "//mediapipe/framework/port:ret_check",
+        "//mediapipe/framework/port:status",
+        "//mediapipe/util:resource_util",
+        "//mediapipe/util/tracking",
+        "//mediapipe/util/tracking:box_detector",
+        "//mediapipe/util/tracking:box_tracker",
+        "//mediapipe/util/tracking:box_tracker_cc_proto",
+        "//mediapipe/util/tracking:flow_packager_cc_proto",
+        "//mediapipe/util/tracking:tracking_visualization_utilities",
+    ] + select({
+        "//mediapipe:android": [
+            "//mediapipe/util/android/file/base",
+        ],
+        "//mediapipe:ios": [
+            "//mediapipe/util/android/file/base",
+        ],
+        "//mediapipe:macos": [
+            "//mediapipe/framework/port:file_helpers",
+        ],
+        "//conditions:default": [
+            "//mediapipe/framework/port:file_helpers",
+        ],
+    }),
+    alwayslink = 1,
+)
+
 cc_library(
    name = "tracked_detection_manager_calculator",
    srcs = ["tracked_detection_manager_calculator.cc"],
    visibility = ["//visibility:public"],
    deps = [
+        ":tracked_detection_manager_calculator_cc_proto",
        "//mediapipe/framework:calculator_framework",
        "//mediapipe/framework/formats:detection_cc_proto",
        "//mediapipe/framework/formats:location_data_cc_proto",
@@ -384,20 +469,18 @@ cc_test(
    ],
 )

-MEDIAPIPE_DEPS = [
-    "//mediapipe/calculators/video:box_tracker_calculator",
-    "//mediapipe/calculators/video:flow_packager_calculator",
-    "//mediapipe/calculators/video:motion_analysis_calculator",
-    "//mediapipe/framework/stream_handler:fixed_size_input_stream_handler",
-    "//mediapipe/framework/stream_handler:sync_set_input_stream_handler",
-]
-
 mediapipe_binary_graph(
    name = "parallel_tracker_binarypb",
    graph = "testdata/parallel_tracker_graph.pbtxt",
    output_name = "testdata/parallel_tracker.binarypb",
    visibility = ["//visibility:public"],
-    deps = MEDIAPIPE_DEPS,
+    deps = [
+        ":box_tracker_calculator",
+        ":flow_packager_calculator",
+        ":motion_analysis_calculator",
+        "//mediapipe/framework/stream_handler:fixed_size_input_stream_handler",
+        "//mediapipe/framework/stream_handler:sync_set_input_stream_handler",
+    ],
 )

 mediapipe_binary_graph(
@@ -405,7 +488,13 @@ mediapipe_binary_graph(
    graph = "testdata/tracker_graph.pbtxt",
    output_name = "testdata/tracker.binarypb",
    visibility = ["//visibility:public"],
-    deps = MEDIAPIPE_DEPS,
+    deps = [
+        ":box_tracker_calculator",
+        ":flow_packager_calculator",
+        ":motion_analysis_calculator",
+        "//mediapipe/framework/stream_handler:fixed_size_input_stream_handler",
+        "//mediapipe/framework/stream_handler:sync_set_input_stream_handler",
+    ],
 )

 cc_test(
@@ -0,0 +1,393 @@
+// Copyright 2019 The MediaPipe Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <stdio.h>
+
+#include <memory>
+#include <unordered_set>
+
+#include "absl/memory/memory.h"
+#include "absl/strings/numbers.h"
+#include "mediapipe/calculators/video/box_detector_calculator.pb.h"
+#include "mediapipe/framework/calculator_framework.h"
+#include "mediapipe/framework/formats/image_frame.h"
+#include "mediapipe/framework/formats/image_frame_opencv.h"
+#include "mediapipe/framework/formats/video_stream_header.h"
+#include "mediapipe/framework/port/integral_types.h"
+#include "mediapipe/framework/port/logging.h"
+#include "mediapipe/framework/port/opencv_core_inc.h"
+#include "mediapipe/framework/port/opencv_features2d_inc.h"
+#include "mediapipe/framework/port/ret_check.h"
+#include "mediapipe/framework/port/status.h"
+#include "mediapipe/util/resource_util.h"
+#include "mediapipe/util/tracking/box_detector.h"
+#include "mediapipe/util/tracking/box_tracker.h"
+#include "mediapipe/util/tracking/box_tracker.pb.h"
+#include "mediapipe/util/tracking/flow_packager.pb.h"
+#include "mediapipe/util/tracking/tracking.h"
+#include "mediapipe/util/tracking/tracking_visualization_utilities.h"
+
+#if defined(MEDIAPIPE_MOBILE)
+#include "mediapipe/util/android/file/base/file.h"
+#include "mediapipe/util/android/file/base/helpers.h"
+#else
+#include "mediapipe/framework/port/file_helpers.h"
+#endif
+
+namespace mediapipe {
+
+// A calculator to detect reappeared box positions from single frame.
+//
+// Input stream:
+//   TRACKING: Input tracking data (proto TrackingData) containing features and
+//             descriptors.
+//   VIDEO:    Optional input video stream tracked boxes are rendered over
+//             (Required if VIZ is specified).
+//   FEATURES: Input feature points (std::vector<cv::KeyPoint>) in the original
+//             pixel space.
+//   DESCRIPTORS: Input feature descriptors (std::vector<float>). Actual feature
+//             dimension needs to be specified in detector_options.
+//   IMAGE_SIZE: Input image dimension.
+//   TRACKED_BOXES : input box tracking result (proto TimedBoxProtoList) from
+//             BoxTrackerCalculator.
+//   ADD_INDEX: Optional std::string containing binary format proto of type
+//             BoxDetectorIndex. Used for adding target index to the detector
+//             search index during runtime.
+//   CANCEL_OBJECT_ID: Optional id of box to be removed. This is recommended
+//             to be used with SyncSetInputStreamHandler.
+//   REACQ_SWITCH: Optional bool for swithcing on and off reacquisition
+//             functionality. User should initialize a graph with box detector
+//             calculator and be able to switch it on and off in runtime.
+//
+// Output streams:
+//   VIZ:   Optional output video stream with rendered box positions
+//          (requires VIDEO to be present)
+//   BOXES: Optional output stream of type TimedBoxProtoList for each lost box.
+//
+// Imput side packets:
+//   INDEX_PROTO_STRING: Optional std::string containing binary format proto of
+//   type
+//                       BoxDetectorIndex. Used for initializing box_detector
+//                       with predefined template images.
+//   FRAME_ALIGNMENT:    Optional integer to indicate alignment_boundary for
+//                       outputing ImageFrame in "VIZ" stream.
+//                       Set to ImageFrame::kDefaultAlignmentBoundary for
+//                       offline pipeline to be compatible with FFmpeg.
+//                       Set to ImageFrame::kGlDefaultAlignmentBoundary for Apps
+//                       to be compatible with GL renderer.
+//   OUTPUT_INDEX_FILENAME: File path to the output index file.
+
+class BoxDetectorCalculator : public CalculatorBase {
+ public:
+  ~BoxDetectorCalculator() override = default;
+
+  static ::mediapipe::Status GetContract(CalculatorContract* cc);
+
+  ::mediapipe::Status Open(CalculatorContext* cc) override;
+  ::mediapipe::Status Process(CalculatorContext* cc) override;
+  ::mediapipe::Status Close(CalculatorContext* cc) override;
+
+ private:
+  BoxDetectorCalculatorOptions options_;
+  std::unique_ptr<BoxDetectorInterface> box_detector_;
+  bool detector_switch_ = true;
+  uint32 frame_alignment_ = ImageFrame::kDefaultAlignmentBoundary;
+  bool write_index_ = false;
+  int box_id_ = 0;
+};
+
+REGISTER_CALCULATOR(BoxDetectorCalculator);
+
+::mediapipe::Status BoxDetectorCalculator::GetContract(CalculatorContract* cc) {
+  if (cc->Inputs().HasTag("TRACKING")) {
+    cc->Inputs().Tag("TRACKING").Set<TrackingData>();
+  }
+
+  if (cc->Inputs().HasTag("TRACKED_BOXES")) {
+    cc->Inputs().Tag("TRACKED_BOXES").Set<TimedBoxProtoList>();
+  }
+
+  if (cc->Inputs().HasTag("VIDEO")) {
+    cc->Inputs().Tag("VIDEO").Set<ImageFrame>();
+  }
+
+  if (cc->Inputs().HasTag("FEATURES")) {
+    RET_CHECK(cc->Inputs().HasTag("DESCRIPTORS"))
+        << "FEATURES and DESCRIPTORS need to be specified together.";
+    cc->Inputs().Tag("FEATURES").Set<std::vector<cv::KeyPoint>>();
+  }
+
+  if (cc->Inputs().HasTag("DESCRIPTORS")) {
+    RET_CHECK(cc->Inputs().HasTag("FEATURES"))
+        << "FEATURES and DESCRIPTORS need to be specified together.";
+    cc->Inputs().Tag("DESCRIPTORS").Set<std::vector<float>>();
+  }
+
+  if (cc->Inputs().HasTag("IMAGE_SIZE")) {
+    cc->Inputs().Tag("IMAGE_SIZE").Set<std::pair<int, int>>();
+  }
+
+  if (cc->Inputs().HasTag("ADD_INDEX")) {
+    cc->Inputs().Tag("ADD_INDEX").Set<std::string>();
+  }
+
+  if (cc->Inputs().HasTag("CANCEL_OBJECT_ID")) {
+    cc->Inputs().Tag("CANCEL_OBJECT_ID").Set<int>();
+  }
+
+  if (cc->Inputs().HasTag("REACQ_SWITCH")) {
+    cc->Inputs().Tag("REACQ_SWITCH").Set<bool>();
+  }
+
+  if (cc->Outputs().HasTag("BOXES")) {
+    cc->Outputs().Tag("BOXES").Set<TimedBoxProtoList>();
+  }
+
+  if (cc->Outputs().HasTag("VIZ")) {
+    RET_CHECK(cc->Inputs().HasTag("VIDEO"))
+        << "Output stream VIZ requires VIDEO to be present.";
+    cc->Outputs().Tag("VIZ").Set<ImageFrame>();
+  }
+
+  if (cc->InputSidePackets().HasTag("INDEX_PROTO_STRING")) {
+    cc->InputSidePackets().Tag("INDEX_PROTO_STRING").Set<std::string>();
+  }
+
+  if (cc->InputSidePackets().HasTag("OUTPUT_INDEX_FILENAME")) {
+    cc->InputSidePackets().Tag("OUTPUT_INDEX_FILENAME").Set<std::string>();
+  }
+
+  if (cc->InputSidePackets().HasTag("FRAME_ALIGNMENT")) {
+    cc->InputSidePackets().Tag("FRAME_ALIGNMENT").Set<int>();
+  }
+
+  return ::mediapipe::OkStatus();
+}
+
+::mediapipe::Status BoxDetectorCalculator::Open(CalculatorContext* cc) {
+  options_ = cc->Options<BoxDetectorCalculatorOptions>();
+  box_detector_ = BoxDetectorInterface::Create(options_.detector_options());
+
+  if (cc->InputSidePackets().HasTag("INDEX_PROTO_STRING")) {
+    BoxDetectorIndex predefined_index;
+    if (!predefined_index.ParseFromString(cc->InputSidePackets()
+                                              .Tag("INDEX_PROTO_STRING")
+                                              .Get<std::string>())) {
+      LOG(FATAL) << "failed to parse BoxDetectorIndex from INDEX_PROTO_STRING";
+    }
+    box_detector_->AddBoxDetectorIndex(predefined_index);
+  }
+
+  for (const auto& filename : options_.index_proto_filename()) {
+    std::string string_path;
+    ASSIGN_OR_RETURN(string_path, PathToResourceAsFile(filename));
+    std::string index_string;
+    MP_RETURN_IF_ERROR(file::GetContents(string_path, &index_string));
+    BoxDetectorIndex predefined_index;
+    if (!predefined_index.ParseFromString(index_string)) {
+      LOG(FATAL)
+          << "failed to parse BoxDetectorIndex from index_proto_filename";
+    }
+    box_detector_->AddBoxDetectorIndex(predefined_index);
+  }
+
+  if (cc->InputSidePackets().HasTag("OUTPUT_INDEX_FILENAME")) {
+    write_index_ = true;
+  }
+
+  if (cc->InputSidePackets().HasTag("FRAME_ALIGNMENT")) {
+    frame_alignment_ = cc->InputSidePackets().Tag("FRAME_ALIGNMENT").Get<int>();
+  }
+
+  return ::mediapipe::OkStatus();
+}
+
+::mediapipe::Status BoxDetectorCalculator::Process(CalculatorContext* cc) {
+  const Timestamp timestamp = cc->InputTimestamp();
+  const int64 timestamp_msec = timestamp.Value() / 1000;
+
+  InputStream* cancel_object_id_stream =
+      cc->Inputs().HasTag("CANCEL_OBJECT_ID")
+          ? &(cc->Inputs().Tag("CANCEL_OBJECT_ID"))
+          : nullptr;
+  if (cancel_object_id_stream && !cancel_object_id_stream->IsEmpty()) {
+    const int cancel_object_id = cancel_object_id_stream->Get<int>();
+    box_detector_->CancelBoxDetection(cancel_object_id);
+  }
+
+  InputStream* add_index_stream = cc->Inputs().HasTag("ADD_INDEX")
+                                      ? &(cc->Inputs().Tag("ADD_INDEX"))
+                                      : nullptr;
+  if (add_index_stream && !add_index_stream->IsEmpty()) {
+    BoxDetectorIndex predefined_index;
+    if (!predefined_index.ParseFromString(
+            add_index_stream->Get<std::string>())) {
+      LOG(FATAL) << "failed to parse BoxDetectorIndex from ADD_INDEX";
+    }
+    box_detector_->AddBoxDetectorIndex(predefined_index);
+  }
+
+  InputStream* reacq_switch_stream = cc->Inputs().HasTag("REACQ_SWITCH")
+                                         ? &(cc->Inputs().Tag("REACQ_SWITCH"))
+                                         : nullptr;
+  if (reacq_switch_stream && !reacq_switch_stream->IsEmpty()) {
+    detector_switch_ = reacq_switch_stream->Get<bool>();
+  }
+
+  if (!detector_switch_) {
+    return ::mediapipe::OkStatus();
+  }
+
+  InputStream* track_stream = cc->Inputs().HasTag("TRACKING")
+                                  ? &(cc->Inputs().Tag("TRACKING"))
+                                  : nullptr;
+  InputStream* video_stream =
+      cc->Inputs().HasTag("VIDEO") ? &(cc->Inputs().Tag("VIDEO")) : nullptr;
+  InputStream* feature_stream = cc->Inputs().HasTag("FEATURES")
+                                    ? &(cc->Inputs().Tag("FEATURES"))
+                                    : nullptr;
+  InputStream* descriptor_stream = cc->Inputs().HasTag("DESCRIPTORS")
+                                       ? &(cc->Inputs().Tag("DESCRIPTORS"))
+                                       : nullptr;
+
+  CHECK(track_stream != nullptr || video_stream != nullptr ||
+        (feature_stream != nullptr && descriptor_stream != nullptr))
+      << "One and only one of {tracking_data, input image frame, "
+         "feature/descriptor} need to be valid.";
+
+  InputStream* tracked_boxes_stream = cc->Inputs().HasTag("TRACKED_BOXES")
+                                          ? &(cc->Inputs().Tag("TRACKED_BOXES"))
+                                          : nullptr;
+  std::unique_ptr<TimedBoxProtoList> detected_boxes(new TimedBoxProtoList());
+
+  if (track_stream != nullptr) {
+    // Detect from tracking data
+    if (track_stream->IsEmpty()) {
+      return ::mediapipe::OkStatus();
+    }
+
+    const TrackingData& tracking_data = track_stream->Get<TrackingData>();
+
+    CHECK(tracked_boxes_stream != nullptr) << "tracked_boxes needed.";
+
+    const TimedBoxProtoList tracked_boxes =
+        tracked_boxes_stream->Get<TimedBoxProtoList>();
+
+    box_detector_->DetectAndAddBox(tracking_data, tracked_boxes, timestamp_msec,
+                                   detected_boxes.get());
+  } else if (video_stream != nullptr) {
+    // Detect from input frame
+    if (video_stream->IsEmpty()) {
+      return ::mediapipe::OkStatus();
+    }
+
+    TimedBoxProtoList tracked_boxes;
+    if (tracked_boxes_stream != nullptr && !tracked_boxes_stream->IsEmpty()) {
+      tracked_boxes = tracked_boxes_stream->Get<TimedBoxProtoList>();
+    }
+
+    // Just directly pass along the image frame data as-is for detection; we
+    // don't need to worry about conforming to a specific alignment here.
+    const cv::Mat input_view =
+        formats::MatView(&video_stream->Get<ImageFrame>());
+    box_detector_->DetectAndAddBox(input_view, tracked_boxes, timestamp_msec,
+                                   detected_boxes.get());
+  } else {
+    if (feature_stream->IsEmpty() || descriptor_stream->IsEmpty()) {
+      return ::mediapipe::OkStatus();
+    }
+
+    const auto& image_size =
+        cc->Inputs().Tag("IMAGE_SIZE").Get<std::pair<int, int>>();
+    float inv_scale = 1.0f / std::max(image_size.first, image_size.second);
+
+    TimedBoxProtoList tracked_boxes;
+    if (tracked_boxes_stream != nullptr && !tracked_boxes_stream->IsEmpty()) {
+      tracked_boxes = tracked_boxes_stream->Get<TimedBoxProtoList>();
+    } else if (write_index_) {
+      auto* box_ptr = tracked_boxes.add_box();
+      box_ptr->set_id(box_id_);
+      box_ptr->set_reacquisition(true);
+      box_ptr->set_aspect_ratio((float)image_size.first /
+                                (float)image_size.second);
+
+      box_ptr->mutable_quad()->add_vertices(0);
+      box_ptr->mutable_quad()->add_vertices(0);
+
+      box_ptr->mutable_quad()->add_vertices(0);
+      box_ptr->mutable_quad()->add_vertices(1);
+
+      box_ptr->mutable_quad()->add_vertices(1);
+      box_ptr->mutable_quad()->add_vertices(1);
+
+      box_ptr->mutable_quad()->add_vertices(1);
+      box_ptr->mutable_quad()->add_vertices(0);
+
+      ++box_id_;
+    }
+
+    const auto& features = feature_stream->Get<std::vector<cv::KeyPoint>>();
+    const int feature_size = features.size();
+    std::vector<Vector2_f> features_vec(feature_size);
+
+    const auto& descriptors = descriptor_stream->Get<std::vector<float>>();
+    const int dims = options_.detector_options().descriptor_dims();
+    CHECK_GE(descriptors.size(), feature_size * dims);
+    cv::Mat descriptors_mat(feature_size, dims, CV_32F);
+    for (int j = 0; j < feature_size; ++j) {
+      features_vec[j].Set(features[j].pt.x * inv_scale,
+                          features[j].pt.y * inv_scale);
+      for (int i = 0; i < dims; ++i) {
+        descriptors_mat.at<float>(j, i) = descriptors[j * dims + i];
+      }
+    }
+
+    box_detector_->DetectAndAddBoxFromFeatures(
+        features_vec, descriptors_mat, tracked_boxes, timestamp_msec,
+        image_size.first * inv_scale, image_size.second * inv_scale,
+        detected_boxes.get());
+  }
+
+  if (cc->Outputs().HasTag("VIZ")) {
+    cv::Mat viz_view;
+    std::unique_ptr<ImageFrame> viz_frame;
+    if (video_stream != nullptr && !video_stream->IsEmpty()) {
+      viz_frame = absl::make_unique<ImageFrame>();
+      viz_frame->CopyFrom(video_stream->Get<ImageFrame>(), frame_alignment_);
+      viz_view = formats::MatView(viz_frame.get());
+    }
+    for (const auto& box : detected_boxes->box()) {
+      RenderBox(box, &viz_view);
+    }
+    cc->Outputs().Tag("VIZ").Add(viz_frame.release(), timestamp);
+  }
+
+  if (cc->Outputs().HasTag("BOXES")) {
+    cc->Outputs().Tag("BOXES").Add(detected_boxes.release(), timestamp);
+  }
+
+  return ::mediapipe::OkStatus();
+}
+
+::mediapipe::Status BoxDetectorCalculator::Close(CalculatorContext* cc) {
+  if (write_index_) {
+    BoxDetectorIndex index = box_detector_->ObtainBoxDetectorIndex();
+    MEDIAPIPE_CHECK_OK(mediapipe::file::SetContents(
+        cc->InputSidePackets().Tag("OUTPUT_INDEX_FILENAME").Get<std::string>(),
+        index.SerializeAsString()));
+  }
+  return ::mediapipe::OkStatus();
+}
+
+}  // namespace mediapipe
@@ -0,0 +1,31 @@
+// Copyright 2019 The MediaPipe Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+syntax = "proto2";
+
+package mediapipe;
+
+import "mediapipe/framework/calculator.proto";
+import "mediapipe/util/tracking/box_detector.proto";
+
+message BoxDetectorCalculatorOptions {
+  extend CalculatorOptions {
+    optional BoxDetectorCalculatorOptions ext = 289746530;
+  }
+
+  optional BoxDetectorOptions detector_options = 1;
+
+  // File path to the template index files.
+  repeated string index_proto_filename = 2;
+}
@@ -18,6 +18,7 @@
 #include <vector>

 #include "absl/container/node_hash_map.h"
+#include "mediapipe/calculators/video/tracked_detection_manager_calculator.pb.h"
 #include "mediapipe/framework/calculator_framework.h"
 #include "mediapipe/framework/formats/detection.pb.h"
 #include "mediapipe/framework/formats/location_data.pb.h"
@@ -139,6 +140,7 @@ Detection GetAxisAlignedDetectionFromTrackedDetection(
 class TrackedDetectionManagerCalculator : public CalculatorBase {
 public:
  static ::mediapipe::Status GetContract(CalculatorContract* cc);
+  ::mediapipe::Status Open(CalculatorContext* cc) override;

  ::mediapipe::Status Process(CalculatorContext* cc) override;

@@ -184,6 +186,15 @@ REGISTER_CALCULATOR(TrackedDetectionManagerCalculator);
  return ::mediapipe::OkStatus();
 }

+::mediapipe::Status TrackedDetectionManagerCalculator::Open(
+    CalculatorContext* cc) {
+  mediapipe::TrackedDetectionManagerCalculatorOptions options =
+      cc->Options<mediapipe::TrackedDetectionManagerCalculatorOptions>();
+  tracked_detection_manager_.SetConfig(
+      options.tracked_detection_manager_options());
+  return ::mediapipe::OkStatus();
+}
+
 ::mediapipe::Status TrackedDetectionManagerCalculator::Process(
    CalculatorContext* cc) {
  if (cc->Inputs().HasTag("TRACKING_BOXES")) {
@@ -0,0 +1,28 @@
+// Copyright 2020 The MediaPipe Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+syntax = "proto2";
+
+package mediapipe;
+
+import "mediapipe/framework/calculator.proto";
+import "mediapipe/util/tracking/tracked_detection_manager_config.proto";
+
+message TrackedDetectionManagerCalculatorOptions {
+  extend CalculatorOptions {
+    optional TrackedDetectionManagerCalculatorOptions ext = 301970230;
+  }
+
+  optional TrackedDetectionManagerConfig tracked_detection_manager_options = 1;
+}
@@ -32,13 +32,14 @@ project.
 2.  Run the Bazel build command to generate the AAR.

    ```bash
-    bazel build -c opt --fat_apk_cpu=arm64-v8a,armeabi-v7a //path/to/the/aar/build/file:aar_name
+    bazel build -c opt --host_crosstool_top=@bazel_tools//tools/cpp:toolchain --fat_apk_cpu=arm64-v8a,armeabi-v7a \
+        //path/to/the/aar/build/file:aar_name
    ```

    For the face detection AAR target we made in the step 1, run:

    ```bash
-    bazel build -c opt --fat_apk_cpu=arm64-v8a,armeabi-v7a \
+    bazel build -c opt --host_crosstool_top=@bazel_tools//tools/cpp:toolchain --fat_apk_cpu=arm64-v8a,armeabi-v7a \
        //mediapipe/examples/android/src/java/com/google/mediapipe/apps/aar_example:mp_face_detection_aar

    # It should print:
@@ -73,6 +73,18 @@ can be easily adapted to run on CPU v.s. GPU.
 *   [Android](./face_detection_mobile_cpu.md)
 *   [iOS](./face_detection_mobile_cpu.md)

+### Face Mesh with GPU
+
+[Face Mesh with GPU](./face_mesh_mobile_gpu.md) illustrates how to run the
+MediaPipe Face Mesh pipeline to perform 3D face landmark estimation in real-time
+on mobile devices, utilizing GPU acceleration. The pipeline is based on
+["Real-time Facial Surface Geometry from Monocular Video on Mobile GPUs"](https://arxiv.org/abs/1907.06724),
+and details of the underlying ML models are described in the
+[model card](https://drive.google.com/file/d/1VFC_wIpw4O7xBOiTgUldl79d9LA-LsnA/view).
+
+*   [Android](./face_mesh_mobile_gpu.md)
+*   [iOS](./face_mesh_mobile_gpu.md)
+
 ### Hand Detection with GPU

 [Hand Detection with GPU](./hand_detection_mobile_gpu.md) illustrates how to use
@@ -84,7 +96,7 @@ MediaPipe with a TFLite model for hand detection in a GPU-accelerated pipeline.
 ### Hand Tracking with GPU

 [Hand Tracking with GPU](./hand_tracking_mobile_gpu.md) illustrates how to use
-MediaPipe with a TFLite model for hand tracking in a GPU-accelerated pipeline.
+MediaPipe with TFLite models for hand tracking in a GPU-accelerated pipeline.

 *   [Android](./hand_tracking_mobile_gpu.md)
 *   [iOS](./hand_tracking_mobile_gpu.md)
@@ -92,7 +104,7 @@ MediaPipe with a TFLite model for hand tracking in a GPU-accelerated pipeline.
 ### 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
+how to use MediaPipe with TFLite models for multi-hand tracking in a
 GPU-accelerated pipeline.

 *   [Android](./multi_hand_tracking_mobile_gpu.md)
@@ -109,6 +121,14 @@ and model details are described in the

 *   [Android](./hair_segmentation_mobile_gpu.md)

+### Template Matching using KNIFT with CPU
+
+[Template Matching using KNIFT on Mobile](./template_matching_mobile_cpu.md)
+shows how to use MediaPipe with TFLite model for template matching using Knift
+on mobile using CPU.
+
+*   [Android](./template_matching_mobile_cpu.md)
+
 ## Desktop

 ### Hello World for C++
@@ -150,11 +170,19 @@ GPU with live video from a webcam.
 *   [Desktop GPU](./face_detection_desktop.md)
 *   [Desktop CPU](./face_detection_desktop.md)

+### Face Mesh on Desktop with Webcam
+
+[Face Mesh on Desktop with Webcam](./face_mesh_desktop.md) shows how to run the
+MediaPipe Face Mesh pipeline to perform 3D face landmark estimation in real-time
+on desktop with webcam input.
+
+*   [Desktop GPU](./face_mesh_desktop.md)
+*   [Desktop CPU](./face_mesh_desktop.md)

 ### Hand Tracking on Desktop with Webcam

 [Hand Tracking on Desktop with Webcam](./hand_tracking_desktop.md) shows how to
-use MediaPipe with a TFLite model for hand tracking on desktop using CPU or GPU
+use MediaPipe with TFLite models for hand tracking on desktop using CPU or GPU
 with live video from a webcam.

 *   [Desktop GPU](./hand_tracking_desktop.md)
@@ -163,8 +191,8 @@ with live video from a webcam.
 ### 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.
+shows how to use MediaPipe with TFLite models 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)
@@ -177,7 +205,7 @@ GPU with live video from a webcam.

 *   [Desktop GPU](./hair_segmentation_desktop.md)

-## Google Coral (machine learning acceleration with Google EdgeTPU)
+## Google Coral (ML acceleration with Google EdgeTPU)

 Below are code samples on how to run MediaPipe on Google Coral Dev Board.

@@ -51,7 +51,7 @@ To build and run the TensorFlow Lite example on desktop (GPU) with Webcam, run:
 ```bash
 # Video from webcam running on desktop GPU
 # This works only for linux currently
-$ bazel build -c opt --copt -DMESA_EGL_NO_X11_HEADERS \
+$ bazel build -c opt --copt -DMESA_EGL_NO_X11_HEADERS --copt -DEGL_NO_X11 \
    mediapipe/examples/desktop/face_detection:face_detection_gpu

 # It should print:
@@ -4,6 +4,8 @@ This doc focuses on the
 [example graph](https://github.com/google/mediapipe/tree/master/mediapipe/graphs/face_detection/face_detection_mobile_cpu.pbtxt)
 that performs face detection with TensorFlow Lite on CPU.

+![face_detection_android_gpu_gif](images/mobile/face_detection_android_gpu.gif)
+
 ## Android

 [Source](https://github.com/google/mediapipe/tree/master/mediapipe/examples/android/src/java/com/google/mediapipe/apps/facedetectioncpu)
@@ -0,0 +1,58 @@
+## Face Mesh on Desktop with Webcam
+
+This doc focuses on running the **MediaPipe Face Mesh** pipeline to perform 3D
+face landmark estimation in real-time on desktop with webcam input. The pipeline
+internally incorporates TensorFlow Lite models. To know more about the models,
+please refer to the model
+[README file](https://github.com/google/mediapipe/tree/master/mediapipe/models/README.md#face-mesh).
+Moreover, if you are interested in running the same pipeline on Android/iOS,
+please see [Face Mesh on Android/iOS](face_mesh_mobile_gpu.md).
+
+-   [Face Mesh on Desktop with Webcam (CPU)](#face-mesh-on-desktop-with-webcam-cpu)
+
+-   [Face Mesh on Desktop with Webcam (GPU)](#face-mesh-on-desktop-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.
+
+### Face Mesh on Desktop with Webcam (CPU)
+
+To build and run Face Mesh on desktop with webcam (CPU), run:
+
+```bash
+$ bazel build -c opt --define MEDIAPIPE_DISABLE_GPU=1 \
+    mediapipe/examples/desktop/face_mesh:face_mesh_cpu
+
+# It should print:
+# Target //mediapipe/examples/desktop/face_mesh:face_mesh_cpu up-to-date:
+#  bazel-bin/mediapipe/examples/desktop/face_mesh/face_mesh_cpu
+
+# This will open up your webcam as long as it is connected. Errors are likely
+# due to your webcam being not accessible.
+$ GLOG_logtostderr=1 bazel-bin/mediapipe/examples/desktop/face_mesh/face_mesh_cpu \
+    --calculator_graph_config_file=mediapipe/graphs/face_mesh/face_mesh_desktop_live.pbtxt
+```
+
+### Face Mesh on Desktop with Webcam (GPU)
+
+Note: please first [check that your GPU is supported](gpu.md#desktop-gpu-linux).
+
+To build and run Face Mesh on desktop with webcam (GPU), run:
+
+```bash
+# This works only for linux currently
+$ bazel build -c opt --copt -DMESA_EGL_NO_X11_HEADERS --copt -DEGL_NO_X11 \
+    mediapipe/examples/desktop/face_mesh:face_mesh_gpu
+
+# It should print:
+# Target //mediapipe/examples/desktop/face_mesh:face_mesh_gpu up-to-date:
+#  bazel-bin/mediapipe/examples/desktop/face_mesh/face_mesh_gpu
+
+# This will open up your webcam as long as it is connected. Errors are likely
+# due to your webcam being not accessible, or GPU drivers not setup properly.
+$ GLOG_logtostderr=1 bazel-bin/mediapipe/examples/desktop/face_mesh/face_mesh_gpu \
+    --calculator_graph_config_file=mediapipe/graphs/face_mesh/face_mesh_desktop_live_gpu.pbtxt
+```
@@ -0,0 +1,90 @@
+# Face Mesh (GPU)
+
+This example focuses on running the **MediaPipe Face Mesh** pipeline on mobile
+devices to perform 3D face landmark estimation in real-time, utilizing GPU
+acceleration. The pipeline internally incorporates TensorFlow Lite models. To
+know more about the models, please refer to the model
+[README file](https://github.com/google/mediapipe/tree/master/mediapipe/models/README.md#face-mesh).
+The pipeline is related to the
+[face detection example](./face_detection_mobile_gpu.md) as it internally
+utilizes face detection and performs landmark estimation only within the
+detected region.
+
+![face_mesh_android_gpu.gif](images/mobile/face_mesh_android_gpu.gif)
+
+**MediaPipe Face Mesh** generates 468 3D face landmarks in real-time on mobile
+devices. In the visualization above, the red dots represent the landmarks, and
+the green lines connecting landmarks illustrate the contours around the eyes,
+eyebrows, lips and the entire face.
+
+## Android
+
+[Source](https://github.com/google/mediapipe/tree/master/mediapipe/examples/android/src/java/com/google/mediapipe/apps/facemeshgpu)
+
+A prebuilt arm64 APK can be
+[downloaded here](https://drive.google.com/open?id=1pUmd7CXCL_onYMbsZo5p91cH0oNnR4gi).
+
+To build the app yourself, run:
+
+```bash
+bazel build -c opt --config=android_arm64 mediapipe/examples/android/src/java/com/google/mediapipe/apps/facemeshgpu
+```
+
+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/facemeshgpu/facemeshgpu.apk
+```
+
+## iOS
+
+[Source](https://github.com/google/mediapipe/tree/master/mediapipe/examples/ios/facemeshgpu).
+
+See the general [instructions](./mediapipe_ios_setup.md) for building iOS
+examples and generating an Xcode project. This will be the FaceMeshGpuApp
+target.
+
+To build on the command line:
+
+```bash
+bazel build -c opt --config=ios_arm64 mediapipe/examples/ios/facemeshgpu:FaceMeshGpuApp
+```
+
+## Graph
+
+The face mesh [main graph](#main-graph) utilizes a
+[face landmark subgraph](#face-landmark-subgraph) from the
+[face landmark module](https://github.com/google/mediapipe/tree/master/mediapipe/modules/face_landmark),
+and renders using a dedicated [face renderer subgraph](#face-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
+
+![face_mesh_mobile_graph](images/mobile/face_mesh_mobile.png)
+
+[Source pbtxt file](https://github.com/google/mediapipe/tree/master/mediapipe/graphs/face_mesh/face_mesh_mobile.pbtxt)
+
+### Face Landmark Subgraph
+
+The
+[face landmark module](https://github.com/google/mediapipe/tree/master/mediapipe/modules/face_landmark)
+contains several subgraphs that can be used to detect and track face landmarks.
+In particular, in this example the
+[FaceLandmarkFrontGPU](https://github.com/google/mediapipe/tree/master/mediapipe/modules/face_landmark/face_landmark_front_gpu.pbtxt)
+subgraph, suitable for images from front-facing cameras (i.e., selfie images)
+and utilizing GPU acceleration, is selected.
+
+![face_landmark_front_gpu_subgraph](images/mobile/face_landmark_front_gpu_subgraph.png)
+
+[Source pbtxt file](https://github.com/google/mediapipe/tree/master/mediapipe/modules/face_landmark/face_landmark_front_gpu.pbtxt)
+
+### Face Renderer Subgraph
+
+![face_renderer_gpu_subgraph](images/mobile/face_renderer_gpu_subgraph.png)
+
+[Source pbtxt file](https://github.com/google/mediapipe/tree/master/mediapipe/graphs/face_mesh/subgraphs/face_renderer_gpu.pbtxt)
@@ -190,10 +190,10 @@ within the MediaPipe framework:
 bazel build --define MEDIAPIPE_DISABLE_GPU=1  <my-target>

 # to enable full GPU support (OpenGL ES 3.1+ & Metal)
-bazel build --copt -DMESA_EGL_NO_X11_HEADERS  <my-target>
+bazel build --copt -DMESA_EGL_NO_X11_HEADERS --copt -DEGL_NO_X11 <my-target>

 # to enable only OpenGL ES 3.0 and below (no GLES 3.1+ features)
-bazel build --copt -DMESA_EGL_NO_X11_HEADERS --copt -DMEDIAPIPE_DISABLE_GL_COMPUTE  <my-target>
+bazel build --copt -DMESA_EGL_NO_X11_HEADERS --copt -DEGL_NO_X11 --copt -DMEDIAPIPE_DISABLE_GL_COMPUTE  <my-target>
 ```

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

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

 # It should print:
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
MediaPipe Teamandchuoling	7bad8fce62	Project import generated by Copybara. GitOrigin-RevId: d0039a576e2db9c0fcefffd26a527df74cbe145b	2020-04-21 22:43:01 -04:00
MediaPipe Teamandjqtang	024f7bf0f1	Project import generated by Copybara. GitOrigin-RevId: a6e7ccd12eb2ad9da2f723eb344658295ed85d46	2020-04-16 20:20:07 -07:00
MediaPipe Teamandchuoling	16e5d7242d	Project import generated by Copybara. GitOrigin-RevId: 72ff4ae24943c2ccf9905bc9e516042b0aa3dd86	2020-04-13 20:15:59 -04:00
MediaPipe Teamandjqtang	4c68eb4a70	Project import generated by Copybara. GitOrigin-RevId: da3cb82c964457b1c719c915a48e545b980af64b	2020-04-06 21:28:19 -07:00
MediaPipe Teamandjqtang	a3d36eee32	Project import generated by Copybara. GitOrigin-RevId: 53a42bf7ad836321123cb7b6c80b0f2e13fbf83e	2020-04-06 19:14:13 -07:00
MediaPipe Teamandjqtang	1722d4b8a2	Project import generated by Copybara. GitOrigin-RevId: 43cd697ec87dcc5cab5051f27960bb77a057399d	2020-03-20 15:28:51 -07:00