Project import generated by Copybara.

GitOrigin-RevId: d91373b4d4d10abef49cab410caa6aadf0875049

mediapipe/calculators/core/BUILD

@@ -165,6 +165,7 @@ cc_library(
     deps = [
         "//mediapipe/framework:calculator_framework",
         "//mediapipe/framework/port:logging",
+        "//mediapipe/framework/port:status",
     ],
     alwayslink = 1,
 )

mediapipe/calculators/core/add_header_calculator.cc

@@ -13,11 +13,12 @@
 // limitations under the License.
 
 #include "mediapipe/framework/calculator_framework.h"
+#include "mediapipe/framework/port/canonical_errors.h"
 #include "mediapipe/framework/port/logging.h"
 
 namespace mediapipe {
 
-// Attach the header from one stream to another stream.
+// Attach the header from a stream or side input to another stream.
 //
 // The header stream (tag HEADER) must not have any packets in it.
 //
@@ -25,17 +26,53 @@ namespace mediapipe {
 // calculator to not need a header or to accept a separate stream with
 // a header, that would be more future proof.
 //
+// Example usage 1:
+// node {
+//   calculator: "AddHeaderCalculator"
+//   input_stream: "DATA:audio"
+//   input_stream: "HEADER:audio_header"
+//   output_stream: "audio_with_header"
+// }
+//
+// Example usage 2:
+// node {
+//   calculator: "AddHeaderCalculator"
+//   input_stream: "DATA:audio"
+//   input_side_packet: "HEADER:audio_header"
+//   output_stream: "audio_with_header"
+// }
+//
 class AddHeaderCalculator : public CalculatorBase {
  public:
   static ::mediapipe::Status GetContract(CalculatorContract* cc) {
-    cc->Inputs().Tag("HEADER").SetNone();
+    bool has_side_input = false;
+    bool has_header_stream = false;
+    if (cc->InputSidePackets().HasTag("HEADER")) {
+      cc->InputSidePackets().Tag("HEADER").SetAny();
+      has_side_input = true;
+    }
+    if (cc->Inputs().HasTag("HEADER")) {
+      cc->Inputs().Tag("HEADER").SetNone();
+      has_header_stream = true;
+    }
+    if (has_side_input == has_header_stream) {
+      return mediapipe::InvalidArgumentError(
+          "Header must be provided via exactly one of side input and input "
+          "stream");
+    }
     cc->Inputs().Tag("DATA").SetAny();
     cc->Outputs().Index(0).SetSameAs(&cc->Inputs().Tag("DATA"));
     return ::mediapipe::OkStatus();
   }
 
   ::mediapipe::Status Open(CalculatorContext* cc) override {
-    const Packet& header = cc->Inputs().Tag("HEADER").Header();
+    Packet header;
+    if (cc->InputSidePackets().HasTag("HEADER")) {
+      header = cc->InputSidePackets().Tag("HEADER");
+    }
+    if (cc->Inputs().HasTag("HEADER")) {
+      header = cc->Inputs().Tag("HEADER").Header();
+    }
     if (!header.IsEmpty()) {
       cc->Outputs().Index(0).SetHeader(header);
     }

mediapipe/calculators/core/add_header_calculator_test.cc

@@ -14,8 +14,10 @@
 
 #include "mediapipe/framework/calculator_framework.h"
 #include "mediapipe/framework/calculator_runner.h"
+#include "mediapipe/framework/port/canonical_errors.h"
 #include "mediapipe/framework/port/gmock.h"
 #include "mediapipe/framework/port/gtest.h"
+#include "mediapipe/framework/port/status.h"
 #include "mediapipe/framework/port/status_matchers.h"
 #include "mediapipe/framework/timestamp.h"
 #include "mediapipe/framework/tool/validate_type.h"
@@ -24,7 +26,7 @@ namespace mediapipe {
 
 class AddHeaderCalculatorTest : public ::testing::Test {};
 
-TEST_F(AddHeaderCalculatorTest, Works) {
+TEST_F(AddHeaderCalculatorTest, HeaderStream) {
   CalculatorGraphConfig::Node node;
   node.set_calculator("AddHeaderCalculator");
   node.add_input_stream("HEADER:header_stream");
@@ -96,4 +98,62 @@ TEST_F(AddHeaderCalculatorTest, NoPacketsOnHeaderStream) {
   ASSERT_FALSE(runner.Run().ok());
 }
 
+TEST_F(AddHeaderCalculatorTest, InputSidePacket) {
+  CalculatorGraphConfig::Node node;
+  node.set_calculator("AddHeaderCalculator");
+  node.add_input_stream("DATA:data_stream");
+  node.add_output_stream("merged_stream");
+  node.add_input_side_packet("HEADER:header");
+
+  CalculatorRunner runner(node);
+
+  // Set header and add 5 packets.
+  runner.MutableSidePackets()->Tag("HEADER") =
+      Adopt(new std::string("my_header"));
+  for (int i = 0; i < 5; ++i) {
+    Packet packet = Adopt(new int(i)).At(Timestamp(i * 1000));
+    runner.MutableInputs()->Tag("DATA").packets.push_back(packet);
+  }
+
+  // Run calculator.
+  MP_ASSERT_OK(runner.Run());
+
+  ASSERT_EQ(1, runner.Outputs().NumEntries());
+
+  // Test output.
+  EXPECT_EQ(std::string("my_header"),
+            runner.Outputs().Index(0).header.Get<std::string>());
+  const std::vector<Packet>& output_packets = runner.Outputs().Index(0).packets;
+  ASSERT_EQ(5, output_packets.size());
+  for (int i = 0; i < 5; ++i) {
+    const int val = output_packets[i].Get<int>();
+    EXPECT_EQ(i, val);
+    EXPECT_EQ(Timestamp(i * 1000), output_packets[i].Timestamp());
+  }
+}
+
+TEST_F(AddHeaderCalculatorTest, UsingBothSideInputAndStream) {
+  CalculatorGraphConfig::Node node;
+  node.set_calculator("AddHeaderCalculator");
+  node.add_input_stream("HEADER:header_stream");
+  node.add_input_stream("DATA:data_stream");
+  node.add_output_stream("merged_stream");
+  node.add_input_side_packet("HEADER:header");
+
+  CalculatorRunner runner(node);
+
+  // Set both headers and add 5 packets.
+  runner.MutableSidePackets()->Tag("HEADER") =
+      Adopt(new std::string("my_header"));
+  runner.MutableSidePackets()->Tag("HEADER") =
+      Adopt(new std::string("my_header"));
+  for (int i = 0; i < 5; ++i) {
+    Packet packet = Adopt(new int(i)).At(Timestamp(i * 1000));
+    runner.MutableInputs()->Tag("DATA").packets.push_back(packet);
+  }
+
+  // Run should fail because header can only be provided one way.
+  EXPECT_EQ(runner.Run().code(), ::mediapipe::InvalidArgumentError("").code());
+}
+
 }  // namespace mediapipe

mediapipe/calculators/core/packet_resampler_calculator.cc

@@ -330,22 +330,27 @@ void PacketResamplerCalculator::UpdateNextOutputTimestampWithJitter() {
     return ::mediapipe::OkStatus();
   }
 
-  LOG_IF(WARNING, frame_time_usec_ <
-                      (cc->InputTimestamp() - last_packet_.Timestamp()).Value())
-      << "Adding jitter is meaningless when upsampling.";
+  if (frame_time_usec_ <
+      (cc->InputTimestamp() - last_packet_.Timestamp()).Value()) {
+    LOG_FIRST_N(WARNING, 2)
+        << "Adding jitter is not very useful when upsampling.";
+  }
 
-  const int64 curr_diff =
-      (next_output_timestamp_ - cc->InputTimestamp()).Value();
-  const int64 last_diff =
-      (next_output_timestamp_ - last_packet_.Timestamp()).Value();
-  if (curr_diff * last_diff > 0) {
-    return ::mediapipe::OkStatus();
+  while (true) {
+    const int64 last_diff =
+        (next_output_timestamp_ - last_packet_.Timestamp()).Value();
+    RET_CHECK_GT(last_diff, 0.0);
+    const int64 curr_diff =
+        (next_output_timestamp_ - cc->InputTimestamp()).Value();
+    if (curr_diff > 0.0) {
+      break;
+    }
+    OutputWithinLimits(cc, (std::abs(curr_diff) > last_diff
+                                ? last_packet_
+                                : cc->Inputs().Get(input_data_id_).Value())
+                               .At(next_output_timestamp_));
+    UpdateNextOutputTimestampWithJitter();
   }
-  OutputWithinLimits(cc, (std::abs(curr_diff) > std::abs(last_diff)
-                              ? last_packet_
-                              : cc->Inputs().Get(input_data_id_).Value())
-                             .At(next_output_timestamp_));
-  UpdateNextOutputTimestampWithJitter();
   return ::mediapipe::OkStatus();
 }
 

mediapipe/calculators/image/image_cropping_calculator.cc

@@ -501,8 +501,11 @@ void ImageCroppingCalculator::GetOutputDimensions(CalculatorContext* cc,
     row_max = std::max(row_max, transformed_points_[i * 2 + 1]);
   }
 
-  *dst_width = std::round((col_max - col_min) * src_width);
-  *dst_height = std::round((row_max - row_min) * src_height);
+  int width = static_cast<int>(std::round((col_max - col_min) * src_width));
+  int height = static_cast<int>(std::round((row_max - row_min) * src_height));
+  // Minimum output dimension 1x1 prevents creation of textures with 0x0.
+  *dst_width = std::max(1, width);
+  *dst_height = std::max(1, height);
 }
 
 }  // namespace mediapipe

mediapipe/calculators/util/BUILD

@@ -983,3 +983,31 @@ cc_test(
         "//mediapipe/framework/port:parse_text_proto",
     ],
 )
+
+cc_library(
+    name = "detections_to_timed_box_list_calculator",
+    srcs = ["detections_to_timed_box_list_calculator.cc"],
+    visibility = ["//visibility:public"],
+    deps = [
+        "//mediapipe/framework:calculator_framework",
+        "//mediapipe/framework/formats:detection_cc_proto",
+        "//mediapipe/framework/formats:location_data_cc_proto",
+        "//mediapipe/framework/port:ret_check",
+        "//mediapipe/framework/port:status",
+        "//mediapipe/util/tracking:box_tracker",
+    ],
+    alwayslink = 1,
+)
+
+cc_library(
+    name = "detection_unique_id_calculator",
+    srcs = ["detection_unique_id_calculator.cc"],
+    visibility = ["//visibility:public"],
+    deps = [
+        "//mediapipe/framework:calculator_framework",
+        "//mediapipe/framework/formats:detection_cc_proto",
+        "//mediapipe/framework/port:ret_check",
+        "//mediapipe/framework/port:status",
+    ],
+    alwayslink = 1,
+)

mediapipe/calculators/util/detection_unique_id_calculator.cc

@@ -0,0 +1,110 @@
+// Copyright 2019 The MediaPipe Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "mediapipe/framework/calculator_framework.h"
+#include "mediapipe/framework/formats/detection.pb.h"
+#include "mediapipe/framework/port/status.h"
+
+namespace mediapipe {
+
+namespace {
+
+constexpr char kDetectionsTag[] = "DETECTIONS";
+constexpr char kDetectionListTag[] = "DETECTION_LIST";
+
+// Each detection processed by DetectionUniqueIDCalculator will be assigned an
+// unique id that starts from 1. If a detection already has an ID other than 0,
+// the ID will be overwritten.
+static int64 detection_id = 0;
+
+inline int GetNextDetectionId() { return ++detection_id; }
+
+}  // namespace
+
+// Assign a unique id to detections.
+// Note that the calculator will consume the input vector of Detection or
+// DetectionList. So the input stream can not be connected to other calculators.
+//
+// Example config:
+// node {
+//   calculator: "DetectionUniqueIdCalculator"
+//   input_stream: "DETECTIONS:detections"
+//   output_stream: "DETECTIONS:output_detections"
+// }
+class DetectionUniqueIdCalculator : public CalculatorBase {
+ public:
+  static ::mediapipe::Status GetContract(CalculatorContract* cc) {
+    RET_CHECK(cc->Inputs().HasTag(kDetectionListTag) ||
+              cc->Inputs().HasTag(kDetectionsTag))
+        << "None of the input streams are provided.";
+
+    if (cc->Inputs().HasTag(kDetectionListTag)) {
+      RET_CHECK(cc->Outputs().HasTag(kDetectionListTag));
+      cc->Inputs().Tag(kDetectionListTag).Set<DetectionList>();
+      cc->Outputs().Tag(kDetectionListTag).Set<DetectionList>();
+    }
+    if (cc->Inputs().HasTag(kDetectionsTag)) {
+      RET_CHECK(cc->Outputs().HasTag(kDetectionsTag));
+      cc->Inputs().Tag(kDetectionsTag).Set<std::vector<Detection>>();
+      cc->Outputs().Tag(kDetectionsTag).Set<std::vector<Detection>>();
+    }
+
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Open(CalculatorContext* cc) override {
+    cc->SetOffset(::mediapipe::TimestampDiff(0));
+    return ::mediapipe::OkStatus();
+  }
+  ::mediapipe::Status Process(CalculatorContext* cc) override;
+};
+REGISTER_CALCULATOR(DetectionUniqueIdCalculator);
+
+::mediapipe::Status DetectionUniqueIdCalculator::Process(
+    CalculatorContext* cc) {
+  if (cc->Inputs().HasTag(kDetectionListTag) &&
+      !cc->Inputs().Tag(kDetectionListTag).IsEmpty()) {
+    auto result =
+        cc->Inputs().Tag(kDetectionListTag).Value().Consume<DetectionList>();
+    if (result.ok()) {
+      auto detection_list = std::move(result).ValueOrDie();
+      for (Detection& detection : *detection_list->mutable_detection()) {
+        detection.set_detection_id(GetNextDetectionId());
+      }
+      cc->Outputs()
+          .Tag(kDetectionListTag)
+          .Add(detection_list.release(), cc->InputTimestamp());
+    }
+  }
+
+  if (cc->Inputs().HasTag(kDetectionsTag) &&
+      !cc->Inputs().Tag(kDetectionsTag).IsEmpty()) {
+    auto result = cc->Inputs()
+                      .Tag(kDetectionsTag)
+                      .Value()
+                      .Consume<std::vector<Detection>>();
+    if (result.ok()) {
+      auto detections = std::move(result).ValueOrDie();
+      for (Detection& detection : *detections) {
+        detection.set_detection_id(GetNextDetectionId());
+      }
+      cc->Outputs()
+          .Tag(kDetectionsTag)
+          .Add(detections.release(), cc->InputTimestamp());
+    }
+  }
+  return ::mediapipe::OkStatus();
+}
+
+}  // namespace mediapipe

mediapipe/calculators/util/detections_to_render_data_calculator.cc

@@ -39,6 +39,8 @@ constexpr char kKeypointLabel[] = "KEYPOINT";
 // The ratio of detection label font height to the height of detection bounding
 // box.
 constexpr double kLabelToBoundingBoxRatio = 0.1;
+// Perserve 2 decimal digits.
+constexpr float kNumScoreDecimalDigitsMultipler = 100;
 
 }  // namespace
 
@@ -235,18 +237,26 @@ void DetectionsToRenderDataCalculator::AddLabels(
     std::string label_str = detection.label().empty()
                                 ? absl::StrCat(detection.label_id(i))
                                 : detection.label(i);
+    const float rounded_score =
+        std::round(detection.score(i) * kNumScoreDecimalDigitsMultipler) /
+        kNumScoreDecimalDigitsMultipler;
     std::string label_and_score =
-        absl::StrCat(label_str, options.text_delimiter(), detection.score(i),
+        absl::StrCat(label_str, options.text_delimiter(), rounded_score,
                      options.text_delimiter());
     label_and_scores.push_back(label_and_score);
   }
   std::vector<std::string> labels;
+  if (options.render_detection_id()) {
+    const std::string detection_id_str =
+        absl::StrCat("Id: ", detection.detection_id());
+    labels.push_back(detection_id_str);
+  }
   if (options.one_label_per_line()) {
-    labels.swap(label_and_scores);
+    labels.insert(labels.end(), label_and_scores.begin(),
+                  label_and_scores.end());
   } else {
     labels.push_back(absl::StrJoin(label_and_scores, ""));
   }
-
   // Add the render annotations for "label(_id),score".
   for (int i = 0; i < labels.size(); ++i) {
     auto label = labels.at(i);

mediapipe/calculators/util/detections_to_render_data_calculator.proto

@@ -53,4 +53,7 @@ message DetectionsToRenderDataCalculatorOptions {
   // instances of this calculator are present in the graph, this value
   // should be unique among them.
   optional string scene_class = 7 [default = "DETECTION"];
+
+  // If true, renders the detection id in the first line before the labels.
+  optional bool render_detection_id = 8 [default = false];
 }

mediapipe/calculators/util/detections_to_timed_box_list_calculator.cc

@@ -0,0 +1,110 @@
+// Copyright 2019 The MediaPipe Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "mediapipe/framework/calculator_framework.h"
+#include "mediapipe/framework/formats/detection.pb.h"
+#include "mediapipe/framework/formats/location_data.pb.h"
+#include "mediapipe/framework/port/status.h"
+#include "mediapipe/util/tracking/box_tracker.h"
+
+namespace mediapipe {
+
+namespace {
+
+constexpr char kDetectionsTag[] = "DETECTIONS";
+constexpr char kDetectionListTag[] = "DETECTION_LIST";
+constexpr char kBoxesTag[] = "BOXES";
+
+}  // namespace
+
+// A calculator that converts Detection proto to TimedBoxList proto for
+// tracking.
+//
+// Please note that only Location Data formats of RELATIVE_BOUNDING_BOX are
+// supported.
+//
+// Example config:
+// node {
+//   calculator: "DetectionsToTimedBoxListCalculator"
+//   input_stream: "DETECTIONS:detections"
+//   output_stream: "BOXES:boxes"
+// }
+class DetectionsToTimedBoxListCalculator : public CalculatorBase {
+ public:
+  static ::mediapipe::Status GetContract(CalculatorContract* cc) {
+    RET_CHECK(cc->Inputs().HasTag(kDetectionListTag) ||
+              cc->Inputs().HasTag(kDetectionsTag))
+        << "None of the input streams are provided.";
+    if (cc->Inputs().HasTag(kDetectionListTag)) {
+      cc->Inputs().Tag(kDetectionListTag).Set<DetectionList>();
+    }
+    if (cc->Inputs().HasTag(kDetectionsTag)) {
+      cc->Inputs().Tag(kDetectionsTag).Set<std::vector<Detection>>();
+    }
+    cc->Outputs().Tag(kBoxesTag).Set<TimedBoxProtoList>();
+    return ::mediapipe::OkStatus();
+  }
+
+  ::mediapipe::Status Open(CalculatorContext* cc) override {
+    cc->SetOffset(TimestampDiff(0));
+    return ::mediapipe::OkStatus();
+  }
+  ::mediapipe::Status Process(CalculatorContext* cc) override;
+
+ private:
+  void ConvertDetectionToTimedBox(const Detection& detection,
+                                  TimedBoxProto* box, CalculatorContext* cc);
+};
+REGISTER_CALCULATOR(DetectionsToTimedBoxListCalculator);
+
+::mediapipe::Status DetectionsToTimedBoxListCalculator::Process(
+    CalculatorContext* cc) {
+  auto output_timed_box_list = absl::make_unique<TimedBoxProtoList>();
+
+  if (cc->Inputs().HasTag(kDetectionListTag)) {
+    const auto& detection_list =
+        cc->Inputs().Tag(kDetectionListTag).Get<DetectionList>();
+    for (const auto& detection : detection_list.detection()) {
+      TimedBoxProto* box = output_timed_box_list->add_box();
+      ConvertDetectionToTimedBox(detection, box, cc);
+    }
+  }
+  if (cc->Inputs().HasTag(kDetectionsTag)) {
+    const auto& detections =
+        cc->Inputs().Tag(kDetectionsTag).Get<std::vector<Detection>>();
+    for (const auto& detection : detections) {
+      TimedBoxProto* box = output_timed_box_list->add_box();
+      ConvertDetectionToTimedBox(detection, box, cc);
+    }
+  }
+
+  cc->Outputs().Tag(kBoxesTag).Add(output_timed_box_list.release(),
+                                   cc->InputTimestamp());
+  return ::mediapipe::OkStatus();
+}
+
+void DetectionsToTimedBoxListCalculator::ConvertDetectionToTimedBox(
+    const Detection& detection, TimedBoxProto* box, CalculatorContext* cc) {
+  const auto& relative_bounding_box =
+      detection.location_data().relative_bounding_box();
+  box->set_left(relative_bounding_box.xmin());
+  box->set_right(relative_bounding_box.xmin() + relative_bounding_box.width());
+  box->set_top(relative_bounding_box.ymin());
+  box->set_bottom(relative_bounding_box.ymin() +
+                  relative_bounding_box.height());
+  box->set_id(detection.detection_id());
+  box->set_time_msec(cc->InputTimestamp().Microseconds() / 1000);
+}
+
+}  // namespace mediapipe

mediapipe/calculators/video/BUILD

@@ -37,6 +37,67 @@ proto_library(
     deps = ["//mediapipe/framework:calculator_proto"],
 )
 
+proto_library(
+    name = "motion_analysis_calculator_proto",
+    srcs = ["motion_analysis_calculator.proto"],
+    deps = [
+        "//mediapipe/framework:calculator_proto",
+        "//mediapipe/util/tracking:motion_analysis_proto",
+    ],
+)
+
+proto_library(
+    name = "flow_packager_calculator_proto",
+    srcs = ["flow_packager_calculator.proto"],
+    deps = [
+        "//mediapipe/framework:calculator_proto",
+        "//mediapipe/util/tracking:flow_packager_proto",
+    ],
+)
+
+proto_library(
+    name = "box_tracker_calculator_proto",
+    srcs = ["box_tracker_calculator.proto"],
+    visibility = ["//visibility:public"],
+    deps = [
+        "//mediapipe/framework:calculator_proto",
+        "//mediapipe/util/tracking:box_tracker_proto",
+    ],
+)
+
+mediapipe_cc_proto_library(
+    name = "motion_analysis_calculator_cc_proto",
+    srcs = ["motion_analysis_calculator.proto"],
+    cc_deps = [
+        "//mediapipe/framework:calculator_cc_proto",
+        "//mediapipe/util/tracking:motion_analysis_cc_proto",
+    ],
+    visibility = ["//visibility:public"],
+    deps = [":motion_analysis_calculator_proto"],
+)
+
+mediapipe_cc_proto_library(
+    name = "flow_packager_calculator_cc_proto",
+    srcs = ["flow_packager_calculator.proto"],
+    cc_deps = [
+        "//mediapipe/framework:calculator_cc_proto",
+        "//mediapipe/util/tracking:flow_packager_cc_proto",
+    ],
+    visibility = ["//visibility:public"],
+    deps = [":flow_packager_calculator_proto"],
+)
+
+mediapipe_cc_proto_library(
+    name = "box_tracker_calculator_cc_proto",
+    srcs = ["box_tracker_calculator.proto"],
+    cc_deps = [
+        "//mediapipe/framework:calculator_cc_proto",
+        "//mediapipe/util/tracking:box_tracker_cc_proto",
+    ],
+    visibility = ["//visibility:public"],
+    deps = [":box_tracker_calculator_proto"],
+)
+
 mediapipe_cc_proto_library(
     name = "flow_to_image_calculator_cc_proto",
     srcs = ["flow_to_image_calculator.proto"],
@@ -131,6 +192,94 @@ cc_library(
     alwayslink = 1,
 )
 
+cc_library(
+    name = "motion_analysis_calculator",
+    srcs = ["motion_analysis_calculator.cc"],
+    visibility = ["//visibility:public"],
+    deps = [
+        ":motion_analysis_calculator_cc_proto",
+        "//mediapipe/framework:calculator_framework",
+        "//mediapipe/framework/formats:image_frame",
+        "//mediapipe/framework/formats:image_frame_opencv",
+        "//mediapipe/framework/formats:video_stream_header",
+        "//mediapipe/framework/port:integral_types",
+        "//mediapipe/framework/port:logging",
+        "//mediapipe/framework/port:ret_check",
+        "//mediapipe/framework/port:status",
+        "//mediapipe/util/tracking:camera_motion",
+        "//mediapipe/util/tracking:camera_motion_cc_proto",
+        "//mediapipe/util/tracking:frame_selection_cc_proto",
+        "//mediapipe/util/tracking:motion_analysis",
+        "//mediapipe/util/tracking:motion_estimation",
+        "//mediapipe/util/tracking:motion_models",
+        "//mediapipe/util/tracking:region_flow_cc_proto",
+        "@com_google_absl//absl/strings",
+    ],
+    alwayslink = 1,
+)
+
+cc_library(
+    name = "flow_packager_calculator",
+    srcs = ["flow_packager_calculator.cc"],
+    visibility = ["//visibility:public"],
+    deps = [
+        ":flow_packager_calculator_cc_proto",
+        "//mediapipe/framework:calculator_framework",
+        "//mediapipe/framework/port:integral_types",
+        "//mediapipe/framework/port:logging",
+        "//mediapipe/util/tracking:camera_motion_cc_proto",
+        "//mediapipe/util/tracking:flow_packager",
+        "//mediapipe/util/tracking:region_flow_cc_proto",
+        "@com_google_absl//absl/strings",
+        "@com_google_absl//absl/strings:str_format",
+    ],
+    alwayslink = 1,
+)
+
+cc_library(
+    name = "box_tracker_calculator",
+    srcs = ["box_tracker_calculator.cc"],
+    visibility = ["//visibility:public"],
+    deps = [
+        ":box_tracker_calculator_cc_proto",
+        "//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:parse_text_proto",
+        "//mediapipe/framework/port:ret_check",
+        "//mediapipe/framework/port:status",
+        "//mediapipe/framework/tool:options_util",
+        "//mediapipe/util/tracking",
+        "//mediapipe/util/tracking:box_tracker",
+        "//mediapipe/util/tracking:tracking_visualization_utilities",
+        "@com_google_absl//absl/strings",
+    ],
+    alwayslink = 1,
+)
+
+cc_library(
+    name = "tracked_detection_manager_calculator",
+    srcs = ["tracked_detection_manager_calculator.cc"],
+    visibility = ["//visibility:public"],
+    deps = [
+        "//mediapipe/framework:calculator_framework",
+        "//mediapipe/framework/formats:detection_cc_proto",
+        "//mediapipe/framework/formats:location_data_cc_proto",
+        "//mediapipe/framework/formats:rect_cc_proto",
+        "//mediapipe/framework/port:status",
+        "//mediapipe/util/tracking",
+        "//mediapipe/util/tracking:box_tracker",
+        "//mediapipe/util/tracking:tracked_detection",
+        "//mediapipe/util/tracking:tracked_detection_manager",
+        "//mediapipe/util/tracking:tracking_visualization_utilities",
+        "@com_google_absl//absl/container:node_hash_map",
+    ],
+    alwayslink = 1,
+)
+
 filegroup(
     name = "test_videos",
     srcs = [
@@ -201,3 +350,64 @@ cc_test(
         "//mediapipe/framework/port:parse_text_proto",
     ],
 )
+
+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,
+)
+
+mediapipe_binary_graph(
+    name = "tracker_binarypb",
+    graph = "testdata/tracker_graph.pbtxt",
+    output_name = "testdata/tracker.binarypb",
+    visibility = ["//visibility:public"],
+    deps = MEDIAPIPE_DEPS,
+)
+
+cc_test(
+    name = "tracking_graph_test",
+    size = "small",
+    srcs = ["tracking_graph_test.cc"],
+    copts = ["-DPARALLEL_INVOKER_ACTIVE"] + select({
+        "//mediapipe:apple": [],
+        "//mediapipe:android": [],
+        "//conditions:default": [],
+    }),
+    data = [
+        ":testdata/lenna.png",
+        ":testdata/parallel_tracker.binarypb",
+        ":testdata/tracker.binarypb",
+    ],
+    deps = [
+        ":box_tracker_calculator",
+        ":box_tracker_calculator_cc_proto",
+        ":flow_packager_calculator",
+        ":motion_analysis_calculator",
+        "//mediapipe/framework:calculator_cc_proto",
+        "//mediapipe/framework:calculator_framework",
+        "//mediapipe/framework:packet",
+        "//mediapipe/framework/deps:file_path",
+        "//mediapipe/framework/formats:image_frame",
+        "//mediapipe/framework/port:advanced_proto",
+        "//mediapipe/framework/port:core_proto",
+        "//mediapipe/framework/port:file_helpers",
+        "//mediapipe/framework/port:gtest_main",
+        "//mediapipe/framework/port:opencv_highgui",
+        "//mediapipe/framework/port:status",
+        "//mediapipe/framework/stream_handler:fixed_size_input_stream_handler",
+        "//mediapipe/framework/stream_handler:sync_set_input_stream_handler",
+        "//mediapipe/util/tracking:box_tracker_cc_proto",
+        "//mediapipe/util/tracking:tracking_cc_proto",
+    ],
+)

mediapipe/calculators/video/box_tracker_calculator.proto

@@ -0,0 +1,55 @@
+// 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_tracker.proto";
+
+message BoxTrackerCalculatorOptions {
+  extend CalculatorOptions {
+    optional BoxTrackerCalculatorOptions ext = 268767860;
+  }
+
+  optional BoxTrackerOptions tracker_options = 1;
+
+  // Initial position to be tracked. Can also be supplied as side packet or
+  // as input stream.
+  optional TimedBoxProtoList initial_position = 2;
+
+  // If set and VIZ stream is present, renders tracking data into the
+  // visualization.
+  optional bool visualize_tracking_data = 3 [default = false];
+
+  // If set and VIZ stream is present, renders the box state
+  // into the visualization.
+  optional bool visualize_state = 4 [default = false];
+
+  // If set and VIZ stream is present, renders the internal box state
+  // into the visualization.
+  optional bool visualize_internal_state = 5 [default = false];
+
+  // Size of the track data cache during streaming mode. This allows to buffer
+  // track_data's for fast forward tracking, i.e. any TimedBox received
+  // via input stream START_POS can be tracked towards the current track head
+  // (i.e. last received TrackingData). Measured in number of frames.
+  optional int32 streaming_track_data_cache_size = 6 [default = 0];
+
+  // Add a transition period of N frames to smooth the jump from original
+  // tracking to reset start pos with motion compensation. The transition will
+  // be a linear decay of original tracking result. 0 means no transition.
+  optional int32 start_pos_transition_frames = 7 [default = 0];
+}

mediapipe/calculators/video/flow_packager_calculator.cc

@@ -0,0 +1,281 @@
+// 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 <fstream>
+#include <memory>
+
+#include "absl/strings/str_format.h"
+#include "absl/strings/string_view.h"
+#include "mediapipe/calculators/video/flow_packager_calculator.pb.h"
+#include "mediapipe/framework/calculator_framework.h"
+#include "mediapipe/framework/port/integral_types.h"
+#include "mediapipe/framework/port/logging.h"
+#include "mediapipe/util/tracking/camera_motion.pb.h"
+#include "mediapipe/util/tracking/flow_packager.h"
+#include "mediapipe/util/tracking/region_flow.pb.h"
+
+namespace mediapipe {
+
+using mediapipe::CameraMotion;
+using mediapipe::FlowPackager;
+using mediapipe::RegionFlowFeatureList;
+using mediapipe::TrackingData;
+using mediapipe::TrackingDataChunk;
+
+// A calculator that packages input CameraMotion and RegionFlowFeatureList
+// into a TrackingData and optionally writes TrackingDataChunks to file.
+//
+// Input stream:
+//   FLOW:       Input region flow (proto RegionFlowFeatureList).
+//   CAMERA:     Input camera stream (proto CameraMotion, optional).
+//
+// Input side packets:
+//   CACHE_DIR:  Optional caching directory tracking files are written to.
+//
+// Output streams.
+//   TRACKING:       Output tracking data (proto TrackingData, per frame
+//                   optional).
+//   TRACKING_CHUNK: Output tracking chunks (proto TrackingDataChunk,
+//                   per chunk, optional), output at the first timestamp
+//                   of each chunk.
+//   COMPLETE:       Optional output packet sent on PreStream to
+//                   to signal downstream calculators that all data has been
+//                   processed and calculator is closed. Can be used to indicate
+//                   that all data as been written to CACHE_DIR.
+class FlowPackagerCalculator : public CalculatorBase {
+ public:
+  ~FlowPackagerCalculator() 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;
+
+  // Writes passed chunk to disk.
+  void WriteChunk(const TrackingDataChunk& chunk) const;
+
+  // Initializes next chunk for tracking beginning from last frame of
+  // current chunk (Chunking is design with one frame overlap).
+  void PrepareCurrentForNextChunk(TrackingDataChunk* chunk);
+
+ private:
+  FlowPackagerCalculatorOptions options_;
+
+  // Caching options.
+  bool use_caching_ = false;
+  bool build_chunk_ = false;
+  std::string cache_dir_;
+  int chunk_idx_ = -1;
+  TrackingDataChunk tracking_chunk_;
+
+  int frame_idx_ = 0;
+
+  Timestamp prev_timestamp_;
+  std::unique_ptr<FlowPackager> flow_packager_;
+};
+
+REGISTER_CALCULATOR(FlowPackagerCalculator);
+
+::mediapipe::Status FlowPackagerCalculator::GetContract(
+    CalculatorContract* cc) {
+  if (!cc->Inputs().HasTag("FLOW")) {
+    return tool::StatusFail("No input flow was specified.");
+  }
+
+  cc->Inputs().Tag("FLOW").Set<RegionFlowFeatureList>();
+
+  if (cc->Inputs().HasTag("CAMERA")) {
+    cc->Inputs().Tag("CAMERA").Set<CameraMotion>();
+  }
+  if (cc->Outputs().HasTag("TRACKING")) {
+    cc->Outputs().Tag("TRACKING").Set<TrackingData>();
+  }
+  if (cc->Outputs().HasTag("TRACKING_CHUNK")) {
+    cc->Outputs().Tag("TRACKING_CHUNK").Set<TrackingDataChunk>();
+  }
+  if (cc->Outputs().HasTag("COMPLETE")) {
+    cc->Outputs().Tag("COMPLETE").Set<bool>();
+  }
+
+  if (cc->InputSidePackets().HasTag("CACHE_DIR")) {
+    cc->InputSidePackets().Tag("CACHE_DIR").Set<std::string>();
+  }
+
+  return ::mediapipe::OkStatus();
+}
+
+::mediapipe::Status FlowPackagerCalculator::Open(CalculatorContext* cc) {
+  options_ = cc->Options<FlowPackagerCalculatorOptions>();
+
+  flow_packager_.reset(new FlowPackager(options_.flow_packager_options()));
+
+  use_caching_ = cc->InputSidePackets().HasTag("CACHE_DIR");
+  build_chunk_ = use_caching_ || cc->Outputs().HasTag("TRACKING_CHUNK");
+  if (use_caching_) {
+    cache_dir_ = cc->InputSidePackets().Tag("CACHE_DIR").Get<std::string>();
+  }
+
+  return ::mediapipe::OkStatus();
+}
+
+::mediapipe::Status FlowPackagerCalculator::Process(CalculatorContext* cc) {
+  InputStream* flow_stream = &(cc->Inputs().Tag("FLOW"));
+  const RegionFlowFeatureList& flow = flow_stream->Get<RegionFlowFeatureList>();
+
+  const Timestamp timestamp = flow_stream->Value().Timestamp();
+
+  const CameraMotion* camera_motion = nullptr;
+  if (cc->Inputs().HasTag("CAMERA")) {
+    InputStream* camera_stream = &(cc->Inputs().Tag("CAMERA"));
+    camera_motion = &camera_stream->Get<CameraMotion>();
+  }
+
+  std::unique_ptr<TrackingData> tracking_data(new TrackingData());
+
+  flow_packager_->PackFlow(flow, camera_motion, tracking_data.get());
+
+  if (build_chunk_) {
+    if (chunk_idx_ < 0) {  // Lazy init, determine first start.
+      chunk_idx_ =
+          timestamp.Value() / 1000 / options_.caching_chunk_size_msec();
+      tracking_chunk_.set_first_chunk(true);
+    }
+    CHECK_GE(chunk_idx_, 0);
+
+    TrackingDataChunk::Item* item = tracking_chunk_.add_item();
+    item->set_frame_idx(frame_idx_);
+    item->set_timestamp_usec(timestamp.Value());
+    if (frame_idx_ > 0) {
+      item->set_prev_timestamp_usec(prev_timestamp_.Value());
+    }
+    if (cc->Outputs().HasTag("TRACKING")) {
+      // Need to copy as output is requested.
+      *item->mutable_tracking_data() = *tracking_data;
+    } else {
+      item->mutable_tracking_data()->Swap(tracking_data.get());
+    }
+
+    const int next_chunk_msec =
+        options_.caching_chunk_size_msec() * (chunk_idx_ + 1);
+
+    if (timestamp.Value() / 1000 >= next_chunk_msec) {
+      if (cc->Outputs().HasTag("TRACKING_CHUNK")) {
+        cc->Outputs()
+            .Tag("TRACKING_CHUNK")
+            .Add(new TrackingDataChunk(tracking_chunk_),
+                 Timestamp(tracking_chunk_.item(0).timestamp_usec()));
+      }
+      if (use_caching_) {
+        WriteChunk(tracking_chunk_);
+      }
+      PrepareCurrentForNextChunk(&tracking_chunk_);
+    }
+  }
+
+  if (cc->Outputs().HasTag("TRACKING")) {
+    cc->Outputs()
+        .Tag("TRACKING")
+        .Add(tracking_data.release(), flow_stream->Value().Timestamp());
+  }
+
+  prev_timestamp_ = timestamp;
+  ++frame_idx_;
+  return ::mediapipe::OkStatus();
+}
+
+::mediapipe::Status FlowPackagerCalculator::Close(CalculatorContext* cc) {
+  if (frame_idx_ > 0) {
+    tracking_chunk_.set_last_chunk(true);
+    if (cc->Outputs().HasTag("TRACKING_CHUNK")) {
+      cc->Outputs()
+          .Tag("TRACKING_CHUNK")
+          .Add(new TrackingDataChunk(tracking_chunk_),
+               Timestamp(tracking_chunk_.item(0).timestamp_usec()));
+    }
+
+    if (use_caching_) {
+      WriteChunk(tracking_chunk_);
+    }
+  }
+
+  if (cc->Outputs().HasTag("COMPLETE")) {
+    cc->Outputs().Tag("COMPLETE").Add(new bool(true), Timestamp::PreStream());
+  }
+
+  return ::mediapipe::OkStatus();
+}
+
+void FlowPackagerCalculator::WriteChunk(const TrackingDataChunk& chunk) const {
+  if (chunk.item_size() == 0) {
+    LOG(ERROR) << "Write chunk called with empty tracking data."
+               << "This can only occur if the spacing between frames "
+               << "is larger than the requested chunk size. Try increasing "
+               << "the chunk size";
+    return;
+  }
+
+  auto format_runtime =
+      absl::ParsedFormat<'d'>::New(options_.cache_file_format());
+
+  std::string chunk_file;
+  if (format_runtime) {
+    chunk_file =
+        cache_dir_ + "/" + absl::StrFormat(*format_runtime, chunk_idx_);
+  } else {
+    LOG(ERROR) << "chache_file_format wrong. fall back to chunk_%04d.";
+    chunk_file = cache_dir_ + "/" + absl::StrFormat("chunk_%04d", chunk_idx_);
+  }
+
+  std::string data;
+  chunk.SerializeToString(&data);
+
+  const char* temp_filename = tempnam(cache_dir_.c_str(), nullptr);
+  std::ofstream out_file(temp_filename);
+  if (!out_file) {
+    LOG(ERROR) << "Could not open " << temp_filename;
+  } else {
+    out_file.write(data.data(), data.size());
+  }
+
+  if (rename(temp_filename, chunk_file.c_str()) != 0) {
+    LOG(ERROR) << "Failed to rename to " << chunk_file;
+  }
+
+  LOG(INFO) << "Wrote chunk : " << chunk_file;
+}
+
+void FlowPackagerCalculator::PrepareCurrentForNextChunk(
+    TrackingDataChunk* chunk) {
+  CHECK(chunk);
+  if (chunk->item_size() == 0) {
+    LOG(ERROR) << "Called with empty chunk. Unexpected.";
+    return;
+  }
+
+  chunk->set_first_chunk(false);
+
+  // Buffer last item for next chunk.
+  TrackingDataChunk::Item last_item;
+  last_item.Swap(chunk->mutable_item(chunk->item_size() - 1));
+
+  chunk->Clear();
+  chunk->add_item()->Swap(&last_item);
+
+  ++chunk_idx_;
+}
+
+}  // namespace mediapipe

mediapipe/calculators/video/flow_packager_calculator.proto

@@ -0,0 +1,36 @@
+// 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/flow_packager.proto";
+
+message FlowPackagerCalculatorOptions {
+  extend CalculatorOptions {
+    optional FlowPackagerCalculatorOptions ext = 271236147;
+  }
+
+  optional mediapipe.FlowPackagerOptions flow_packager_options = 1;
+
+  // Chunk size for caching files that are written to the externally specified
+  // caching directory. Specified in msec.
+  // Note that each chunk always contains at its end the first frame of the
+  // next chunk (to enable forward tracking across chunk boundaries).
+  optional int32 caching_chunk_size_msec = 2 [default = 2500];
+
+  optional string cache_file_format = 3 [default = "chunk_%04d"];
+}

mediapipe/calculators/video/motion_analysis_calculator.cc

@@ -0,0 +1,988 @@
+// Copyright 2019 The MediaPipe Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <cmath>
+#include <fstream>
+#include <memory>
+
+#include "absl/strings/numbers.h"
+#include "absl/strings/str_split.h"
+#include "absl/strings/string_view.h"
+#include "mediapipe/calculators/video/motion_analysis_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/ret_check.h"
+#include "mediapipe/framework/port/status.h"
+#include "mediapipe/util/tracking/camera_motion.h"
+#include "mediapipe/util/tracking/camera_motion.pb.h"
+#include "mediapipe/util/tracking/frame_selection.pb.h"
+#include "mediapipe/util/tracking/motion_analysis.h"
+#include "mediapipe/util/tracking/motion_estimation.h"
+#include "mediapipe/util/tracking/motion_models.h"
+#include "mediapipe/util/tracking/region_flow.pb.h"
+
+namespace mediapipe {
+
+using mediapipe::AffineAdapter;
+using mediapipe::CameraMotion;
+using mediapipe::FrameSelectionResult;
+using mediapipe::Homography;
+using mediapipe::HomographyAdapter;
+using mediapipe::LinearSimilarityModel;
+using mediapipe::MixtureHomography;
+using mediapipe::MixtureRowWeights;
+using mediapipe::MotionAnalysis;
+using mediapipe::ProjectViaFit;
+using mediapipe::RegionFlowComputationOptions;
+using mediapipe::RegionFlowFeatureList;
+using mediapipe::SalientPointFrame;
+using mediapipe::TranslationModel;
+
+const char kOptionsTag[] = "OPTIONS";
+
+// A calculator that performs motion analysis on an incoming video stream.
+//
+// Input streams:  (at least one of them is required).
+//   VIDEO:     The input video stream (ImageFrame, sRGB, sRGBA or GRAY8).
+//   SELECTION: Optional input stream to perform analysis only on selected
+//              frames. If present needs to contain camera motion
+//              and features.
+//
+// Input side packets:
+//   CSV_FILE:  Read motion models as homographies from CSV file. Expected
+//              to be defined in the frame domain (un-normalized).
+//              Should store 9 floats per row.
+//              Specify number of homographies per frames via option
+//              meta_models_per_frame. For values > 1, MixtureHomographies
+//              are created, for value == 1, a single Homography is used.
+//   DOWNSAMPLE: Optionally specify downsampling factor via input side packet
+//               overriding value in the graph settings.
+// Output streams (all are optional).
+//   FLOW:      Sparse feature tracks in form of proto RegionFlowFeatureList.
+//   CAMERA:    Camera motion as proto CameraMotion describing the per frame-
+//              pair motion. Has VideoHeader from input video.
+//   SALIENCY:  Foreground saliency (objects moving different from the
+//              background) as proto SalientPointFrame.
+//   VIZ:       Visualization stream as ImageFrame, sRGB, visualizing
+//              features and saliency (set via
+//              analysis_options().visualization_options())
+//   DENSE_FG:  Dense foreground stream, describing per-pixel foreground-
+//              ness as confidence between 0 (background) and 255
+//              (foreground). Output is ImageFrame (GRAY8).
+//   VIDEO_OUT: Optional output stream when SELECTION is used. Output is input
+//              VIDEO at the selected frames. Required VIDEO to be present.
+//   GRAY_VIDEO_OUT: Optional output stream for downsampled, grayscale video.
+//                   Requires VIDEO to be present and SELECTION to not be used.
+class MotionAnalysisCalculator : public CalculatorBase {
+  // TODO: Activate once leakr approval is ready.
+  // typedef com::google::android::libraries::micro::proto::Data HomographyData;
+
+ public:
+  ~MotionAnalysisCalculator() 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:
+  // Outputs results to Outputs() if MotionAnalysis buffered sufficient results.
+  // Otherwise no-op. Set flush to true to force output of all buffered data.
+  void OutputMotionAnalyzedFrames(bool flush, CalculatorContext* cc);
+
+  // Lazy init function to be called on Process.
+  ::mediapipe::Status InitOnProcess(InputStream* video_stream,
+                                    InputStream* selection_stream);
+
+  // Parses CSV file contents to homographies.
+  bool ParseModelCSV(const std::string& contents,
+                     std::deque<Homography>* homographies);
+
+  // Turns list of 9-tuple floating values into set of homographies.
+  bool HomographiesFromValues(const std::vector<float>& homog_values,
+                              std::deque<Homography>* homographies);
+
+  // Appends CameraMotions and features from homographies.
+  // Set append_identity to true to add an identity transform to the beginning
+  // of the each list *in addition* to the motions derived from homographies.
+  void AppendCameraMotionsFromHomographies(
+      const std::deque<Homography>& homographies, bool append_identity,
+      std::deque<CameraMotion>* camera_motions,
+      std::deque<RegionFlowFeatureList>* features);
+
+  // Helper function to subtract current metadata motion from features. Used
+  // for hybrid estimation case.
+  void SubtractMetaMotion(const CameraMotion& meta_motion,
+                          RegionFlowFeatureList* features);
+
+  // Inverse of above function to add back meta motion and replace
+  // feature location with originals after estimation.
+  void AddMetaMotion(const CameraMotion& meta_motion,
+                     const RegionFlowFeatureList& meta_features,
+                     RegionFlowFeatureList* features, CameraMotion* motion);
+
+  MotionAnalysisCalculatorOptions options_;
+  int frame_width_ = -1;
+  int frame_height_ = -1;
+  int frame_idx_ = 0;
+
+  // Buffers incoming video frame packets (if visualization output is requested)
+  std::vector<Packet> packet_buffer_;
+
+  // Buffers incoming timestamps until MotionAnalysis is ready to output via
+  // above OutputMotionAnalyzedFrames.
+  std::vector<Timestamp> timestamp_buffer_;
+
+  // Input indicators for each stream.
+  bool selection_input_ = false;
+  bool video_input_ = false;
+
+  // Output indicators for each stream.
+  bool region_flow_feature_output_ = false;
+  bool camera_motion_output_ = false;
+  bool saliency_output_ = false;
+  bool visualize_output_ = false;
+  bool dense_foreground_output_ = false;
+  bool video_output_ = false;
+  bool grayscale_output_ = false;
+  bool csv_file_input_ = false;
+
+  // Inidicates if saliency should be computed.
+  bool with_saliency_ = false;
+
+  // Set if hybrid meta analysis - see proto for details.
+  bool hybrid_meta_analysis_ = false;
+
+  // Concatenated motions for each selected frame. Used in case
+  // hybrid estimation is requested to fallback to valid models.
+  std::deque<CameraMotion> selected_motions_;
+
+  // Normalized homographies from CSV file or metadata.
+  std::deque<Homography> meta_homographies_;
+  std::deque<CameraMotion> meta_motions_;
+  std::deque<RegionFlowFeatureList> meta_features_;
+
+  // Offset into above meta_motions_ and features_ when using
+  // hybrid meta analysis.
+  int hybrid_meta_offset_ = 0;
+
+  std::unique_ptr<MotionAnalysis> motion_analysis_;
+
+  std::unique_ptr<MixtureRowWeights> row_weights_;
+};
+
+REGISTER_CALCULATOR(MotionAnalysisCalculator);
+
+::mediapipe::Status MotionAnalysisCalculator::GetContract(
+    CalculatorContract* cc) {
+  if (cc->Inputs().HasTag("VIDEO")) {
+    cc->Inputs().Tag("VIDEO").Set<ImageFrame>();
+  }
+
+  // Optional input stream from frame selection calculator.
+  if (cc->Inputs().HasTag("SELECTION")) {
+    cc->Inputs().Tag("SELECTION").Set<FrameSelectionResult>();
+  }
+
+  RET_CHECK(cc->Inputs().HasTag("VIDEO") || cc->Inputs().HasTag("SELECTION"))
+      << "Either VIDEO, SELECTION must be specified.";
+
+  if (cc->Outputs().HasTag("FLOW")) {
+    cc->Outputs().Tag("FLOW").Set<RegionFlowFeatureList>();
+  }
+
+  if (cc->Outputs().HasTag("CAMERA")) {
+    cc->Outputs().Tag("CAMERA").Set<CameraMotion>();
+  }
+
+  if (cc->Outputs().HasTag("SALIENCY")) {
+    cc->Outputs().Tag("SALIENCY").Set<SalientPointFrame>();
+  }
+
+  if (cc->Outputs().HasTag("VIZ")) {
+    cc->Outputs().Tag("VIZ").Set<ImageFrame>();
+  }
+
+  if (cc->Outputs().HasTag("DENSE_FG")) {
+    cc->Outputs().Tag("DENSE_FG").Set<ImageFrame>();
+  }
+
+  if (cc->Outputs().HasTag("VIDEO_OUT")) {
+    cc->Outputs().Tag("VIDEO_OUT").Set<ImageFrame>();
+  }
+
+  if (cc->Outputs().HasTag("GRAY_VIDEO_OUT")) {
+    // We only output grayscale video if we're actually performing full region-
+    // flow analysis on the video.
+    RET_CHECK(cc->Inputs().HasTag("VIDEO") &&
+              !cc->Inputs().HasTag("SELECTION"));
+    cc->Outputs().Tag("GRAY_VIDEO_OUT").Set<ImageFrame>();
+  }
+
+  if (cc->InputSidePackets().HasTag("CSV_FILE")) {
+    cc->InputSidePackets().Tag("CSV_FILE").Set<std::string>();
+  }
+  if (cc->InputSidePackets().HasTag("DOWNSAMPLE")) {
+    cc->InputSidePackets().Tag("DOWNSAMPLE").Set<float>();
+  }
+
+  if (cc->InputSidePackets().HasTag(kOptionsTag)) {
+    cc->InputSidePackets().Tag(kOptionsTag).Set<CalculatorOptions>();
+  }
+
+  return ::mediapipe::OkStatus();
+}
+
+::mediapipe::Status MotionAnalysisCalculator::Open(CalculatorContext* cc) {
+  options_ =
+      tool::RetrieveOptions(cc->Options<MotionAnalysisCalculatorOptions>(),
+                            cc->InputSidePackets(), kOptionsTag);
+
+  video_input_ = cc->Inputs().HasTag("VIDEO");
+  selection_input_ = cc->Inputs().HasTag("SELECTION");
+  region_flow_feature_output_ = cc->Outputs().HasTag("FLOW");
+  camera_motion_output_ = cc->Outputs().HasTag("CAMERA");
+  saliency_output_ = cc->Outputs().HasTag("SALIENCY");
+  visualize_output_ = cc->Outputs().HasTag("VIZ");
+  dense_foreground_output_ = cc->Outputs().HasTag("DENSE_FG");
+  video_output_ = cc->Outputs().HasTag("VIDEO_OUT");
+  grayscale_output_ = cc->Outputs().HasTag("GRAY_VIDEO_OUT");
+  csv_file_input_ = cc->InputSidePackets().HasTag("CSV_FILE");
+  hybrid_meta_analysis_ = options_.meta_analysis() ==
+                          MotionAnalysisCalculatorOptions::META_ANALYSIS_HYBRID;
+
+  if (video_output_) {
+    RET_CHECK(selection_input_) << "VIDEO_OUT requires SELECTION input";
+  }
+
+  if (selection_input_) {
+    switch (options_.selection_analysis()) {
+      case MotionAnalysisCalculatorOptions::NO_ANALYSIS_USE_SELECTION:
+        RET_CHECK(!visualize_output_)
+            << "Visualization not supported for NO_ANALYSIS_USE_SELECTION";
+        RET_CHECK(!dense_foreground_output_)
+            << "Dense foreground not supported for NO_ANALYSIS_USE_SELECTION";
+        RET_CHECK(!saliency_output_)
+            << "Saliency output not supported for NO_ANALYSIS_USE_SELECTION";
+        break;
+
+      case MotionAnalysisCalculatorOptions::ANALYSIS_RECOMPUTE:
+      case MotionAnalysisCalculatorOptions::ANALYSIS_WITH_SEED:
+        RET_CHECK(video_input_) << "Need video input for feature tracking.";
+        break;
+
+      case MotionAnalysisCalculatorOptions::ANALYSIS_FROM_FEATURES:
+        // Nothing to add here.
+        break;
+    }
+  }
+
+  if (visualize_output_ || dense_foreground_output_ || video_output_) {
+    RET_CHECK(video_input_) << "Video input required.";
+  }
+
+  if (csv_file_input_) {
+    RET_CHECK(!selection_input_)
+        << "Can not use selection input with csv input.";
+    if (!hybrid_meta_analysis_) {
+      RET_CHECK(!saliency_output_ && !visualize_output_ &&
+                !dense_foreground_output_ && !grayscale_output_)
+          << "CSV file and meta input only supports flow and camera motion "
+          << "output when using metadata only.";
+    }
+  }
+
+  if (csv_file_input_) {
+    // Read from file and parse.
+    const std::string filename =
+        cc->InputSidePackets().Tag("CSV_FILE").Get<std::string>();
+
+    std::string file_contents;
+    std::ifstream input_file(filename, std::ios::in);
+    input_file.seekg(0, std::ios::end);
+    const int file_length = input_file.tellg();
+    file_contents.resize(file_length);
+    input_file.seekg(0, std::ios::beg);
+    input_file.read(&file_contents[0], file_length);
+    input_file.close();
+
+    RET_CHECK(ParseModelCSV(file_contents, &meta_homographies_))
+        << "Could not parse CSV file";
+  }
+
+  // Get video header from video or selection input if present.
+  const VideoHeader* video_header = nullptr;
+  if (video_input_ && !cc->Inputs().Tag("VIDEO").Header().IsEmpty()) {
+    video_header = &(cc->Inputs().Tag("VIDEO").Header().Get<VideoHeader>());
+  } else if (selection_input_ &&
+             !cc->Inputs().Tag("SELECTION").Header().IsEmpty()) {
+    video_header = &(cc->Inputs().Tag("SELECTION").Header().Get<VideoHeader>());
+  } else {
+    LOG(WARNING) << "No input video header found. Downstream calculators "
+                    "expecting video headers are likely to fail.";
+  }
+
+  with_saliency_ = options_.analysis_options().compute_motion_saliency();
+  // Force computation of saliency if requested as output.
+  if (cc->Outputs().HasTag("SALIENCY")) {
+    with_saliency_ = true;
+    if (!options_.analysis_options().compute_motion_saliency()) {
+      LOG(WARNING) << "Enable saliency computation. Set "
+                   << "compute_motion_saliency to true to silence this "
+                   << "warning.";
+      options_.mutable_analysis_options()->set_compute_motion_saliency(true);
+    }
+  }
+
+  if (options_.bypass_mode()) {
+    cc->SetOffset(TimestampDiff(0));
+  }
+
+  if (cc->InputSidePackets().HasTag("DOWNSAMPLE")) {
+    options_.mutable_analysis_options()
+        ->mutable_flow_options()
+        ->set_downsample_factor(
+            cc->InputSidePackets().Tag("DOWNSAMPLE").Get<float>());
+  }
+
+  // If no video header is provided, just return and initialize on the first
+  // Process() call.
+  if (video_header == nullptr) {
+    return ::mediapipe::OkStatus();
+  }
+
+  ////////////// EARLY RETURN; ONLY HEADER OUTPUT SHOULD GO HERE ///////////////
+
+  if (visualize_output_) {
+    cc->Outputs().Tag("VIZ").SetHeader(Adopt(new VideoHeader(*video_header)));
+  }
+
+  if (video_output_) {
+    cc->Outputs()
+        .Tag("VIDEO_OUT")
+        .SetHeader(Adopt(new VideoHeader(*video_header)));
+  }
+
+  if (cc->Outputs().HasTag("DENSE_FG")) {
+    std::unique_ptr<VideoHeader> foreground_header(
+        new VideoHeader(*video_header));
+    foreground_header->format = ImageFormat::GRAY8;
+    cc->Outputs().Tag("DENSE_FG").SetHeader(Adopt(foreground_header.release()));
+  }
+
+  if (cc->Outputs().HasTag("CAMERA")) {
+    cc->Outputs().Tag("CAMERA").SetHeader(
+        Adopt(new VideoHeader(*video_header)));
+  }
+
+  if (cc->Outputs().HasTag("SALIENCY")) {
+    cc->Outputs()
+        .Tag("SALIENCY")
+        .SetHeader(Adopt(new VideoHeader(*video_header)));
+  }
+
+  return ::mediapipe::OkStatus();
+}
+
+::mediapipe::Status MotionAnalysisCalculator::Process(CalculatorContext* cc) {
+  if (options_.bypass_mode()) {
+    return ::mediapipe::OkStatus();
+  }
+
+  InputStream* video_stream =
+      video_input_ ? &(cc->Inputs().Tag("VIDEO")) : nullptr;
+  InputStream* selection_stream =
+      selection_input_ ? &(cc->Inputs().Tag("SELECTION")) : nullptr;
+
+  // Checked on Open.
+  CHECK(video_stream || selection_stream);
+
+  // Lazy init.
+  if (frame_width_ < 0 || frame_height_ < 0) {
+    MP_RETURN_IF_ERROR(InitOnProcess(video_stream, selection_stream));
+  }
+
+  const Timestamp timestamp = cc->InputTimestamp();
+  if ((csv_file_input_) && !hybrid_meta_analysis_) {
+    if (camera_motion_output_) {
+      RET_CHECK(!meta_motions_.empty()) << "Insufficient metadata.";
+
+      CameraMotion output_motion = meta_motions_.front();
+      meta_motions_.pop_front();
+      output_motion.set_timestamp_usec(timestamp.Value());
+      cc->Outputs().Tag("CAMERA").Add(new CameraMotion(output_motion),
+                                      timestamp);
+    }
+
+    if (region_flow_feature_output_) {
+      RET_CHECK(!meta_features_.empty()) << "Insufficient frames in CSV file";
+      RegionFlowFeatureList output_features = meta_features_.front();
+      meta_features_.pop_front();
+
+      output_features.set_timestamp_usec(timestamp.Value());
+      cc->Outputs().Tag("FLOW").Add(new RegionFlowFeatureList(output_features),
+                                    timestamp);
+    }
+
+    ++frame_idx_;
+    return ::mediapipe::OkStatus();
+  }
+
+  if (motion_analysis_ == nullptr) {
+    // We do not need MotionAnalysis when using just metadata.
+    motion_analysis_.reset(new MotionAnalysis(options_.analysis_options(),
+                                              frame_width_, frame_height_));
+  }
+
+  std::unique_ptr<FrameSelectionResult> frame_selection_result;
+  // Always use frame if selection is not activated.
+  bool use_frame = !selection_input_;
+  if (selection_input_) {
+    CHECK(selection_stream);
+
+    // Fill in timestamps we process.
+    if (!selection_stream->Value().IsEmpty()) {
+      ASSIGN_OR_RETURN(
+          frame_selection_result,
+          selection_stream->Value().ConsumeOrCopy<FrameSelectionResult>());
+      use_frame = true;
+
+      // Make sure both features and camera motion are present.
+      RET_CHECK(frame_selection_result->has_camera_motion() &&
+                frame_selection_result->has_features())
+          << "Frame selection input error at: " << timestamp
+          << " both camera motion and features need to be "
+             "present in FrameSelectionResult. "
+          << frame_selection_result->has_camera_motion() << " , "
+          << frame_selection_result->has_features();
+    }
+  }
+
+  if (selection_input_ && use_frame &&
+      options_.selection_analysis() ==
+          MotionAnalysisCalculatorOptions::NO_ANALYSIS_USE_SELECTION) {
+    // Output concatenated results, nothing to compute here.
+    if (camera_motion_output_) {
+      cc->Outputs().Tag("CAMERA").Add(
+          frame_selection_result->release_camera_motion(), timestamp);
+    }
+    if (region_flow_feature_output_) {
+      cc->Outputs().Tag("FLOW").Add(frame_selection_result->release_features(),
+                                    timestamp);
+    }
+
+    if (video_output_) {
+      cc->Outputs().Tag("VIDEO_OUT").AddPacket(video_stream->Value());
+    }
+
+    return ::mediapipe::OkStatus();
+  }
+
+  if (use_frame) {
+    if (!selection_input_) {
+      const cv::Mat input_view =
+          formats::MatView(&video_stream->Get<ImageFrame>());
+      if (hybrid_meta_analysis_) {
+        // Seed with meta homography.
+        RET_CHECK(hybrid_meta_offset_ < meta_motions_.size())
+            << "Not enough metadata received for hybrid meta analysis";
+        Homography initial_transform =
+            meta_motions_[hybrid_meta_offset_].homography();
+        std::function<void(RegionFlowFeatureList*)> subtract_helper = std::bind(
+            &MotionAnalysisCalculator::SubtractMetaMotion, this,
+            meta_motions_[hybrid_meta_offset_], std::placeholders::_1);
+
+        // Keep original features before modification around.
+        motion_analysis_->AddFrameGeneric(
+            input_view, timestamp.Value(), initial_transform, nullptr, nullptr,
+            &subtract_helper, &meta_features_[hybrid_meta_offset_]);
+        ++hybrid_meta_offset_;
+      } else {
+        motion_analysis_->AddFrame(input_view, timestamp.Value());
+      }
+    } else {
+      selected_motions_.push_back(frame_selection_result->camera_motion());
+      switch (options_.selection_analysis()) {
+        case MotionAnalysisCalculatorOptions::NO_ANALYSIS_USE_SELECTION:
+          return ::mediapipe::UnknownErrorBuilder(MEDIAPIPE_LOC)
+                 << "Should not reach this point!";
+
+        case MotionAnalysisCalculatorOptions::ANALYSIS_FROM_FEATURES:
+          motion_analysis_->AddFeatures(frame_selection_result->features());
+          break;
+
+        case MotionAnalysisCalculatorOptions::ANALYSIS_RECOMPUTE: {
+          const cv::Mat input_view =
+              formats::MatView(&video_stream->Get<ImageFrame>());
+          motion_analysis_->AddFrame(input_view, timestamp.Value());
+          break;
+        }
+
+        case MotionAnalysisCalculatorOptions::ANALYSIS_WITH_SEED: {
+          Homography homography;
+          CameraMotionToHomography(frame_selection_result->camera_motion(),
+                                   &homography);
+          const cv::Mat input_view =
+              formats::MatView(&video_stream->Get<ImageFrame>());
+          motion_analysis_->AddFrameGeneric(input_view, timestamp.Value(),
+                                            homography, &homography);
+          break;
+        }
+      }
+    }
+
+    timestamp_buffer_.push_back(timestamp);
+    ++frame_idx_;
+
+    VLOG_EVERY_N(0, 100) << "Analyzed frame " << frame_idx_;
+
+    // Buffer input frames only if visualization is requested.
+    if (visualize_output_ || video_output_) {
+      packet_buffer_.push_back(video_stream->Value());
+    }
+
+    // If requested, output grayscale thumbnails
+    if (grayscale_output_) {
+      cv::Mat grayscale_mat = motion_analysis_->GetGrayscaleFrameFromResults();
+      std::unique_ptr<ImageFrame> grayscale_image(new ImageFrame(
+          ImageFormat::GRAY8, grayscale_mat.cols, grayscale_mat.rows));
+      cv::Mat image_frame_mat = formats::MatView(grayscale_image.get());
+      grayscale_mat.copyTo(image_frame_mat);
+
+      cc->Outputs()
+          .Tag("GRAY_VIDEO_OUT")
+          .Add(grayscale_image.release(), timestamp);
+    }
+
+    // Output other results, if we have any yet.
+    OutputMotionAnalyzedFrames(false, cc);
+  }
+
+  return ::mediapipe::OkStatus();
+}
+
+::mediapipe::Status MotionAnalysisCalculator::Close(CalculatorContext* cc) {
+  // Guard against empty videos.
+  if (motion_analysis_) {
+    OutputMotionAnalyzedFrames(true, cc);
+  }
+  if (csv_file_input_) {
+    if (!meta_motions_.empty()) {
+      LOG(ERROR) << "More motions than frames. Unexpected! Remainder: "
+                 << meta_motions_.size();
+    }
+  }
+  return ::mediapipe::OkStatus();
+}
+
+void MotionAnalysisCalculator::OutputMotionAnalyzedFrames(
+    bool flush, CalculatorContext* cc) {
+  std::vector<std::unique_ptr<RegionFlowFeatureList>> features;
+  std::vector<std::unique_ptr<CameraMotion>> camera_motions;
+  std::vector<std::unique_ptr<SalientPointFrame>> saliency;
+
+  const int buffer_size = timestamp_buffer_.size();
+  const int num_results = motion_analysis_->GetResults(
+      flush, &features, &camera_motions, with_saliency_ ? &saliency : nullptr);
+
+  CHECK_LE(num_results, buffer_size);
+
+  if (num_results == 0) {
+    return;
+  }
+
+  for (int k = 0; k < num_results; ++k) {
+    // Region flow features and camera motion for this frame.
+    auto& feature_list = features[k];
+    auto& camera_motion = camera_motions[k];
+    const Timestamp timestamp = timestamp_buffer_[k];
+
+    if (selection_input_ && options_.hybrid_selection_camera()) {
+      if (camera_motion->type() > selected_motions_.front().type()) {
+        // Composited type is more stable.
+        camera_motion->Swap(&selected_motions_.front());
+      }
+      selected_motions_.pop_front();
+    }
+
+    if (hybrid_meta_analysis_) {
+      AddMetaMotion(meta_motions_.front(), meta_features_.front(),
+                    feature_list.get(), camera_motion.get());
+      meta_motions_.pop_front();
+      meta_features_.pop_front();
+    }
+
+    // Video frame for visualization.
+    std::unique_ptr<ImageFrame> visualization_frame;
+    cv::Mat visualization;
+    if (visualize_output_) {
+      // Initialize visualization frame with original frame.
+      visualization_frame.reset(new ImageFrame());
+      visualization_frame->CopyFrom(packet_buffer_[k].Get<ImageFrame>(), 16);
+      visualization = formats::MatView(visualization_frame.get());
+
+      motion_analysis_->RenderResults(
+          *feature_list, *camera_motion,
+          with_saliency_ ? saliency[k].get() : nullptr, &visualization);
+
+      cc->Outputs().Tag("VIZ").Add(visualization_frame.release(), timestamp);
+    }
+
+    // Output dense foreground mask.
+    if (dense_foreground_output_) {
+      std::unique_ptr<ImageFrame> foreground_frame(
+          new ImageFrame(ImageFormat::GRAY8, frame_width_, frame_height_));
+      cv::Mat foreground = formats::MatView(foreground_frame.get());
+      motion_analysis_->ComputeDenseForeground(*feature_list, *camera_motion,
+                                               &foreground);
+      cc->Outputs().Tag("DENSE_FG").Add(foreground_frame.release(), timestamp);
+    }
+
+    // Output flow features if requested.
+    if (region_flow_feature_output_) {
+      cc->Outputs().Tag("FLOW").Add(feature_list.release(), timestamp);
+    }
+
+    // Output camera motion.
+    if (camera_motion_output_) {
+      cc->Outputs().Tag("CAMERA").Add(camera_motion.release(), timestamp);
+    }
+
+    if (video_output_) {
+      cc->Outputs().Tag("VIDEO_OUT").AddPacket(packet_buffer_[k]);
+    }
+
+    // Output saliency.
+    if (saliency_output_) {
+      cc->Outputs().Tag("SALIENCY").Add(saliency[k].release(), timestamp);
+    }
+  }
+
+  if (hybrid_meta_analysis_) {
+    hybrid_meta_offset_ -= num_results;
+    CHECK_GE(hybrid_meta_offset_, 0);
+  }
+
+  timestamp_buffer_.erase(timestamp_buffer_.begin(),
+                          timestamp_buffer_.begin() + num_results);
+
+  if (visualize_output_ || video_output_) {
+    packet_buffer_.erase(packet_buffer_.begin(),
+                         packet_buffer_.begin() + num_results);
+  }
+}
+
+::mediapipe::Status MotionAnalysisCalculator::InitOnProcess(
+    InputStream* video_stream, InputStream* selection_stream) {
+  if (video_stream) {
+    frame_width_ = video_stream->Get<ImageFrame>().Width();
+    frame_height_ = video_stream->Get<ImageFrame>().Height();
+
+    // Ensure image options are set correctly.
+    auto* region_options =
+        options_.mutable_analysis_options()->mutable_flow_options();
+
+    // Use two possible formats to account for different channel orders.
+    RegionFlowComputationOptions::ImageFormat image_format;
+    RegionFlowComputationOptions::ImageFormat image_format2;
+    switch (video_stream->Get<ImageFrame>().Format()) {
+      case ImageFormat::GRAY8:
+        image_format = image_format2 =
+            RegionFlowComputationOptions::FORMAT_GRAYSCALE;
+        break;
+
+      case ImageFormat::SRGB:
+        image_format = RegionFlowComputationOptions::FORMAT_RGB;
+        image_format2 = RegionFlowComputationOptions::FORMAT_BGR;
+        break;
+
+      case ImageFormat::SRGBA:
+        image_format = RegionFlowComputationOptions::FORMAT_RGBA;
+        image_format2 = RegionFlowComputationOptions::FORMAT_BGRA;
+        break;
+
+      default:
+        RET_CHECK(false) << "Unsupported image format.";
+    }
+    if (region_options->image_format() != image_format &&
+        region_options->image_format() != image_format2) {
+      LOG(WARNING) << "Requested image format in RegionFlowComputation "
+                   << "does not match video stream format. Overriding.";
+      region_options->set_image_format(image_format);
+    }
+
+    // Account for downsampling mode INPUT_SIZE. In this case we are handed
+    // already downsampled frames but the resulting CameraMotion should
+    // be computed on higher resolution as specifed by the downsample scale.
+    if (region_options->downsample_mode() ==
+        RegionFlowComputationOptions::DOWNSAMPLE_TO_INPUT_SIZE) {
+      const float scale = region_options->downsample_factor();
+      frame_width_ = static_cast<int>(std::round(frame_width_ * scale));
+      frame_height_ = static_cast<int>(std::round(frame_height_ * scale));
+    }
+  } else if (selection_stream) {
+    const auto& camera_motion =
+        selection_stream->Get<FrameSelectionResult>().camera_motion();
+    frame_width_ = camera_motion.frame_width();
+    frame_height_ = camera_motion.frame_height();
+  } else {
+    LOG(FATAL) << "Either VIDEO or SELECTION stream need to be specified.";
+  }
+
+  // Filled by CSV file parsing.
+  if (!meta_homographies_.empty()) {
+    CHECK(csv_file_input_);
+    AppendCameraMotionsFromHomographies(meta_homographies_,
+                                        true,  // append identity.
+                                        &meta_motions_, &meta_features_);
+    meta_homographies_.clear();
+  }
+
+  // Filter weights before using for hybrid mode.
+  if (hybrid_meta_analysis_) {
+    auto* motion_options =
+        options_.mutable_analysis_options()->mutable_motion_options();
+    motion_options->set_filter_initialized_irls_weights(true);
+  }
+
+  return ::mediapipe::OkStatus();
+}
+
+bool MotionAnalysisCalculator::ParseModelCSV(
+    const std::string& contents, std::deque<Homography>* homographies) {
+  std::vector<absl::string_view> values =
+      absl::StrSplit(contents, absl::ByAnyChar(",\n"));
+
+  // Trim off any empty lines.
+  while (values.back().empty()) {
+    values.pop_back();
+  }
+
+  // Convert to float.
+  std::vector<float> homog_values;
+  homog_values.reserve(values.size());
+
+  for (const auto& value : values) {
+    double value_64f;
+    if (!absl::SimpleAtod(value, &value_64f)) {
+      LOG(ERROR) << "Not a double, expected!";
+      return false;
+    }
+
+    homog_values.push_back(value_64f);
+  }
+
+  return HomographiesFromValues(homog_values, homographies);
+}
+
+bool MotionAnalysisCalculator::HomographiesFromValues(
+    const std::vector<float>& homog_values,
+    std::deque<Homography>* homographies) {
+  CHECK(homographies);
+
+  // Obvious constants are obvious :D
+  constexpr int kHomographyValues = 9;
+  if (homog_values.size() % kHomographyValues != 0) {
+    LOG(ERROR) << "Contents not a multiple of " << kHomographyValues;
+    return false;
+  }
+
+  for (int k = 0; k < homog_values.size(); k += kHomographyValues) {
+    std::vector<double> h_vals(kHomographyValues);
+    for (int l = 0; l < kHomographyValues; ++l) {
+      h_vals[l] = homog_values[k + l];
+    }
+
+    // Normalize last entry to 1.
+    if (h_vals[kHomographyValues - 1] == 0) {
+      LOG(ERROR) << "Degenerate homography, last entry is zero";
+      return false;
+    }
+
+    const double scale = 1.0f / h_vals[kHomographyValues - 1];
+    for (int l = 0; l < kHomographyValues; ++l) {
+      h_vals[l] *= scale;
+    }
+
+    Homography h = HomographyAdapter::FromDoublePointer(h_vals.data(), false);
+    homographies->push_back(h);
+  }
+
+  if (homographies->size() % options_.meta_models_per_frame() != 0) {
+    LOG(ERROR) << "Total homographies not a multiple of specified models "
+               << "per frame.";
+    return false;
+  }
+
+  return true;
+}
+
+void MotionAnalysisCalculator::SubtractMetaMotion(
+    const CameraMotion& meta_motion, RegionFlowFeatureList* features) {
+  if (meta_motion.mixture_homography().model_size() > 0) {
+    CHECK(row_weights_ != nullptr);
+    RegionFlowFeatureListViaTransform(meta_motion.mixture_homography(),
+                                      features, -1.0f,
+                                      1.0f,  // subtract transformed.
+                                      true,  // replace feature loc.
+                                      row_weights_.get());
+  } else {
+    RegionFlowFeatureListViaTransform(meta_motion.homography(), features, -1.0f,
+                                      1.0f,   // subtract transformed.
+                                      true);  // replace feature loc.
+  }
+
+  // Clamp transformed features to domain and handle outliers.
+  const float domain_diam =
+      hypot(features->frame_width(), features->frame_height());
+  const float motion_mag = meta_motion.average_magnitude();
+  // Same irls fraction as used by MODEL_MIXTURE_HOMOGRAPHY scaling in
+  // MotionEstimation.
+  const float irls_fraction = options_.analysis_options()
+                                  .motion_options()
+                                  .irls_mixture_fraction_scale() *
+                              options_.analysis_options()
+                                  .motion_options()
+                                  .irls_motion_magnitude_fraction();
+  float err_scale = std::max(1.0f, motion_mag * irls_fraction);
+
+  const float max_err =
+      options_.meta_outlier_domain_ratio() * domain_diam * err_scale;
+  const float max_err_sq = max_err * max_err;
+
+  for (auto& feature : *features->mutable_feature()) {
+    feature.set_x(
+        std::max(0.0f, std::min(features->frame_width() - 1.0f, feature.x())));
+    feature.set_y(
+        std::max(0.0f, std::min(features->frame_height() - 1.0f, feature.y())));
+    // Label anything with large residual motion an outlier.
+    if (FeatureFlow(feature).Norm2() > max_err_sq) {
+      feature.set_irls_weight(0.0f);
+    }
+  }
+}
+
+void MotionAnalysisCalculator::AddMetaMotion(
+    const CameraMotion& meta_motion, const RegionFlowFeatureList& meta_features,
+    RegionFlowFeatureList* features, CameraMotion* motion) {
+  // Restore old feature location.
+  CHECK_EQ(meta_features.feature_size(), features->feature_size());
+  for (int k = 0; k < meta_features.feature_size(); ++k) {
+    auto feature = features->mutable_feature(k);
+    const auto& meta_feature = meta_features.feature(k);
+    feature->set_x(meta_feature.x());
+    feature->set_y(meta_feature.y());
+    feature->set_dx(meta_feature.dx());
+    feature->set_dy(meta_feature.dy());
+  }
+
+  // Composite camera motion.
+  *motion = ComposeCameraMotion(*motion, meta_motion);
+  // Restore type from metadata, i.e. do not declare motions as invalid.
+  motion->set_type(meta_motion.type());
+  motion->set_match_frame(-1);
+}
+
+void MotionAnalysisCalculator::AppendCameraMotionsFromHomographies(
+    const std::deque<Homography>& homographies, bool append_identity,
+    std::deque<CameraMotion>* camera_motions,
+    std::deque<RegionFlowFeatureList>* features) {
+  CHECK(camera_motions);
+  CHECK(features);
+
+  CameraMotion identity;
+  identity.set_frame_width(frame_width_);
+  identity.set_frame_height(frame_height_);
+
+  *identity.mutable_translation() = TranslationModel();
+  *identity.mutable_linear_similarity() = LinearSimilarityModel();
+  *identity.mutable_homography() = Homography();
+  identity.set_type(CameraMotion::VALID);
+  identity.set_match_frame(0);
+
+  RegionFlowFeatureList empty_list;
+  empty_list.set_long_tracks(true);
+  empty_list.set_match_frame(-1);
+  empty_list.set_frame_width(frame_width_);
+  empty_list.set_frame_height(frame_height_);
+
+  if (append_identity) {
+    camera_motions->push_back(identity);
+    features->push_back(empty_list);
+  }
+
+  const int models_per_frame = options_.meta_models_per_frame();
+  CHECK_GT(models_per_frame, 0) << "At least one model per frame is needed";
+  CHECK_EQ(0, homographies.size() % models_per_frame);
+  const int num_frames = homographies.size() / models_per_frame;
+
+  // Heuristic sigma, similar to what we use for rolling shutter removal.
+  const float mixture_sigma = 1.0f / models_per_frame;
+
+  if (row_weights_ == nullptr) {
+    row_weights_.reset(new MixtureRowWeights(frame_height_,
+                                             frame_height_ / 10,  // 10% margin
+                                             mixture_sigma * frame_height_,
+                                             1.0f, models_per_frame));
+  }
+
+  for (int f = 0; f < num_frames; ++f) {
+    MixtureHomography mix_homog;
+    const int model_start = f * models_per_frame;
+
+    for (int k = 0; k < models_per_frame; ++k) {
+      const Homography& homog = homographies[model_start + k];
+      *mix_homog.add_model() = ModelInvert(homog);
+    }
+
+    CameraMotion c = identity;
+    c.set_match_frame(-1);
+
+    if (mix_homog.model_size() > 1) {
+      *c.mutable_mixture_homography() = mix_homog;
+      c.set_mixture_row_sigma(mixture_sigma);
+
+      for (int k = 0; k < models_per_frame; ++k) {
+        c.add_mixture_inlier_coverage(1.0f);
+      }
+      *c.add_mixture_homography_spectrum() = mix_homog;
+      c.set_rolling_shutter_motion_index(0);
+
+      *c.mutable_homography() = ProjectViaFit<Homography>(
+          mix_homog, frame_width_, frame_height_, row_weights_.get());
+    } else {
+      // Guaranteed to exist because to check that models_per_frame > 0 above.
+      *c.mutable_homography() = mix_homog.model(0);
+    }
+
+    // Project remaining motions down.
+    *c.mutable_linear_similarity() = ProjectViaFit<LinearSimilarityModel>(
+        c.homography(), frame_width_, frame_height_);
+    *c.mutable_translation() = ProjectViaFit<TranslationModel>(
+        c.homography(), frame_width_, frame_height_);
+
+    c.set_average_magnitude(
+        std::hypot(c.translation().dx(), c.translation().dy()));
+
+    camera_motions->push_back(c);
+    features->push_back(empty_list);
+  }
+}
+
+}  // namespace mediapipe

mediapipe/calculators/video/motion_analysis_calculator.proto

@@ -0,0 +1,111 @@
+// 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/motion_analysis.proto";
+
+// Next tag: 10
+message MotionAnalysisCalculatorOptions {
+  extend CalculatorOptions {
+    optional MotionAnalysisCalculatorOptions ext = 270698255;
+  }
+
+  optional mediapipe.MotionAnalysisOptions analysis_options = 1;
+
+  // Determines how optional input SELECTION (if present) is used to compute
+  // the final camera motion.
+  enum SelectionAnalysis {
+    // Recompute camera motion for selected frame neighbors.
+    ANALYSIS_RECOMPUTE = 1;
+
+    // Use composited camera motion and region flow from SELECTION input. No
+    // tracking or re-computation is performed.
+    // Note that in this case only CAMERA, FLOW and VIDEO_OUT tags are
+    // supported as output.
+    NO_ANALYSIS_USE_SELECTION = 2;
+
+    // Recompute camera motion for selected frame neighbors using
+    // features supplied by SELECTION input. No feature tracking is performed.
+    ANALYSIS_FROM_FEATURES = 3;
+
+    // Recomputes camera motion for selected frame neighbors but seeds
+    // initial transform with camera motion from SELECTION input.
+    ANALYSIS_WITH_SEED = 4;
+  }
+
+  optional SelectionAnalysis selection_analysis = 4
+      [default = ANALYSIS_WITH_SEED];
+
+  // If activated when SELECTION input is activated, will replace the computed
+  // camera motion (for any of the ANALYSIS_* case above) with the one supplied
+  // by the frame selection, in case the frame selection one is more stable.
+  // For example, if recomputed camera motion is unstable but the one from
+  // the selection result is stable, will use the stable result instead.
+  optional bool hybrid_selection_camera = 5 [default = false];
+
+  // Determines how optional input META is used to compute the final camera
+  // motion.
+  enum MetaAnalysis {
+    // Uses metadata supplied motions as is.
+    META_ANALYSIS_USE_META = 1;
+
+    // Seeds visual tracking from metadata motions - estimates visual residual
+    // motion and combines with metadata.
+    META_ANALYSIS_HYBRID = 2;
+  }
+
+  optional MetaAnalysis meta_analysis = 8 [default = META_ANALYSIS_USE_META];
+
+  // Determines number of homography models per frame stored in the CSV file
+  // or the homography metadata in META.
+  // For values > 1, MixtureHomographies are created.
+  optional int32 meta_models_per_frame = 6 [default = 1];
+
+  // Used for META_ANALYSIS_HYBRID. Rejects features which flow deviates
+  // domain_ratio * image diagonal size from the ground truth metadata motion.
+  optional float meta_outlier_domain_ratio = 9 [default = 0.0015];
+
+  // If true, the MotionAnalysisCalculator will skip all processing and emit no
+  // packets on any output. This is useful for quickly creating different
+  // versions of a MediaPipe graph without changing its structure, assuming that
+  // downstream calculators can handle missing input packets.
+  // TODO: Remove this hack. See b/36485206 for more details.
+  optional bool bypass_mode = 7 [default = false];
+}
+
+// Taken from
+// java/com/google/android/libraries/microvideo/proto/microvideo.proto to
+// satisfy leakr requirements
+// TODO: Remove and use above proto.
+message HomographyData {
+  // For each frame, there are 12 homography matrices stored. Each matrix is
+  // 3x3 (9 elements). This field will contain 12 x 3 x 3  float values. The
+  // first row of the first homography matrix will be followed by the second row
+  // of the first homography matrix, followed by third row of first homography
+  // matrix, followed by the first row of the second homography matrix, etc.
+  repeated float motion_homography_data = 1 [packed = true];
+
+  // Vector containing histogram counts for individual patches in the frame.
+  repeated uint32 histogram_count_data = 2 [packed = true];
+
+  // The width of the frame at the time metadata was sampled.
+  optional int32 frame_width = 3;
+
+  // The height of the frame at the time metadata was sampled.
+  optional int32 frame_height = 4;
+}

mediapipe/calculators/video/testdata/parallel_tracker_graph.pbtxt

@@ -0,0 +1,134 @@
+input_stream: "image_cpu_frames"
+input_stream: "start_pos"
+input_stream: "ra_track"
+
+num_threads: 4
+
+node: {
+  calculator: "MotionAnalysisCalculator"
+  input_stream: "VIDEO:image_cpu_frames"
+  output_stream: "CAMERA:camera_motion"
+  output_stream: "FLOW:region_flow"
+
+  options: {
+    [mediapipe.MotionAnalysisCalculatorOptions.ext]: {
+      analysis_options: {
+        analysis_policy: ANALYSIS_POLICY_CAMERA_MOBILE
+
+        flow_options: {
+          # Maybe move down to 50
+          fast_estimation_min_block_size: 100
+          top_inlier_sets: 1
+          frac_inlier_error_threshold: 3e-3
+          downsample_mode: DOWNSAMPLE_NONE
+          verification_distance: 5.0
+          verify_long_feature_acceleration: true
+          verify_long_feature_trigger_ratio: 0.1
+          tracking_options: {
+            max_features: 500
+            adaptive_extraction_levels: 2
+            min_eig_val_settings: {
+              adaptive_lowest_quality_level: 2e-4
+            }
+            klt_tracker_implementation: KLT_OPENCV
+          }
+        }
+
+        motion_options: {
+          label_empty_frames_as_valid: false
+        }
+      }
+    }
+  }
+}
+
+node: {
+  calculator: "FlowPackagerCalculator"
+  input_stream: "FLOW:region_flow"
+  input_stream: "CAMERA:camera_motion"
+  output_stream: "TRACKING:tracking_data"
+
+  options: {
+    [mediapipe.FlowPackagerCalculatorOptions.ext]: {
+      flow_packager_options: {
+        binary_tracking_data_support: false
+      }
+    }
+  }
+}
+
+node: {
+  calculator: "BoxTrackerCalculator"
+  input_stream: "TRACKING:tracking_data"
+  input_stream: "START_POS:start_pos"
+  output_stream: "BOXES:boxes"
+  input_side_packet: "OPTIONS:calculator_options"
+
+  input_stream_handler: {
+    input_stream_handler: "SyncSetInputStreamHandler"
+    options: {
+      [mediapipe.SyncSetInputStreamHandlerOptions.ext]: {
+        sync_set: {
+          tag_index: "TRACKING"
+        }
+        sync_set: {
+          tag_index: "START_POS"
+        }
+      }
+    }
+  }
+
+  options: {
+    [mediapipe.BoxTrackerCalculatorOptions.ext]: {
+      tracker_options: {
+        track_step_options: {
+          track_object_and_camera: true
+          tracking_degrees: TRACKING_DEGREE_OBJECT_PERSPECTIVE
+          object_similarity_min_contd_inliers: 6
+          inlier_spring_force: 0.0
+          static_motion_temporal_ratio: 3e-2
+        }
+      }
+      visualize_tracking_data: false
+      streaming_track_data_cache_size: 100
+    }
+  }
+}
+
+node: {
+  calculator: "BoxTrackerCalculator"
+  input_stream: "TRACKING:tracking_data"
+  input_stream: "RA_TRACK:ra_track"
+  output_stream: "RA_BOXES:ra_boxes"
+  input_side_packet: "OPTIONS:calculator_options"
+
+  input_stream_handler: {
+    input_stream_handler: "SyncSetInputStreamHandler"
+    options: {
+      [mediapipe.SyncSetInputStreamHandlerOptions.ext]: {
+        sync_set: {
+          tag_index: "TRACKING"
+        }
+        sync_set: {
+          tag_index: "RA_TRACK"
+        }
+      }
+    }
+  }
+
+  options: {
+    [mediapipe.BoxTrackerCalculatorOptions.ext]: {
+      tracker_options: {
+        track_step_options: {
+          track_object_and_camera: true
+          tracking_degrees: TRACKING_DEGREE_OBJECT_PERSPECTIVE
+          object_similarity_min_contd_inliers: 6
+          inlier_spring_force: 0.0
+          static_motion_temporal_ratio: 3e-2
+        }
+      }
+      visualize_tracking_data: false
+      streaming_track_data_cache_size: 100
+    }
+  }
+}

mediapipe/calculators/video/testdata/tracker_graph.pbtxt

@@ -0,0 +1,122 @@
+input_stream: "image_cpu_frames"
+input_stream: "start_pos"
+input_stream: "cancel_object_id"
+input_stream: "ra_track"
+input_stream: "restart_pos"
+input_stream: "track_time"
+
+num_threads: 4
+
+node: {
+  calculator: "MotionAnalysisCalculator"
+  options: {
+    [mediapipe.MotionAnalysisCalculatorOptions.ext]: {
+      analysis_options: {
+        analysis_policy: ANALYSIS_POLICY_CAMERA_MOBILE
+
+        flow_options: {
+          # Maybe move down to 50
+          fast_estimation_min_block_size: 100
+          top_inlier_sets: 1
+          frac_inlier_error_threshold: 3e-3
+          # For mobile application, downsample before input into graph
+          # and use DOWNSAMPLE_TO_INPUT_SIZE and specify
+          # downsampling_factor option or DOWNSAMPLE input_side_packet
+          downsample_mode: DOWNSAMPLE_TO_INPUT_SIZE
+          verification_distance: 5.0
+          verify_long_feature_acceleration: true
+          verify_long_feature_trigger_ratio: 0.1
+          tracking_options: {
+            max_features: 500
+            corner_extraction_method: EXTRACTION_FAST
+            adaptive_extraction_levels: 2
+            min_eig_val_settings: {
+              adaptive_lowest_quality_level: 2e-4
+            }
+            klt_tracker_implementation: KLT_OPENCV
+          }
+        }
+      }
+    }
+  }
+
+  # Drops packets if calculator cannot keep up with the input rate.
+  input_stream_handler: {
+    input_stream_handler: "FixedSizeInputStreamHandler"
+  }
+
+  input_stream: "VIDEO:image_cpu_frames"
+  input_side_packet: "DOWNSAMPLE:analysis_downsample_factor"
+  input_side_packet: "OPTIONS:calculator_options"
+  output_stream: "CAMERA:camera_motion"
+  output_stream: "FLOW:region_flow"
+}
+
+node: {
+  calculator: "FlowPackagerCalculator"
+
+  input_stream: "FLOW:region_flow"
+  input_stream: "CAMERA:camera_motion"
+  output_stream: "TRACKING:tracking_data"
+  options: {
+    [mediapipe.FlowPackagerCalculatorOptions.ext]: {
+      flow_packager_options: {
+        binary_tracking_data_support: false
+      }
+    }
+  }
+}
+
+node: {
+  calculator: "BoxTrackerCalculator"
+
+  input_side_packet: "OPTIONS:calculator_options"
+  input_stream: "TRACKING:tracking_data"
+  input_stream: "TRACK_TIME:track_time"
+  input_stream: "START_POS:start_pos"
+  input_stream: "RESTART_POS:restart_pos"
+  input_stream: "CANCEL_OBJECT_ID:cancel_object_id"
+  input_stream: "RA_TRACK:ra_track"
+  output_stream: "BOXES:boxes"
+  output_stream: "RA_BOXES:ra_boxes"
+
+  input_stream_handler: {
+    input_stream_handler: "SyncSetInputStreamHandler"
+    options: {
+      [mediapipe.SyncSetInputStreamHandlerOptions.ext]: {
+        sync_set: {
+          tag_index: "TRACKING"
+          tag_index: "TRACK_TIME"
+        }
+        sync_set: {
+          tag_index: "START_POS"
+        }
+        sync_set: {
+          tag_index: "RESTART_POS"
+        }
+        sync_set: {
+          tag_index: "CANCEL_OBJECT_ID"
+        }
+        sync_set: {
+          tag_index: "RA_TRACK"
+        }
+      }
+    }
+  }
+
+  options: {
+    [mediapipe.BoxTrackerCalculatorOptions.ext]: {
+      tracker_options: {
+        track_step_options: {
+          track_object_and_camera: true
+          tracking_degrees: TRACKING_DEGREE_OBJECT_ROTATION_SCALE
+          inlier_spring_force: 0.0
+          static_motion_temporal_ratio: 3e-2
+          object_similarity_min_contd_inliers: 10
+        }
+      }
+      visualize_tracking_data: false
+      streaming_track_data_cache_size: 100
+    }
+  }
+}

mediapipe/calculators/video/tracked_detection_manager_calculator.cc

@@ -0,0 +1,319 @@
+// 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 <unordered_map>
+#include <vector>
+
+#include "absl/container/node_hash_map.h"
+#include "mediapipe/framework/calculator_framework.h"
+#include "mediapipe/framework/formats/detection.pb.h"
+#include "mediapipe/framework/formats/location_data.pb.h"
+#include "mediapipe/framework/formats/rect.pb.h"
+#include "mediapipe/framework/port/status.h"
+#include "mediapipe/util/tracking/box_tracker.h"
+#include "mediapipe/util/tracking/tracked_detection.h"
+#include "mediapipe/util/tracking/tracked_detection_manager.h"
+#include "mediapipe/util/tracking/tracking.h"
+
+namespace mediapipe {
+namespace {
+
+constexpr int kDetectionUpdateTimeOutMS = 5000;
+constexpr char kDetectionsTag[] = "DETECTIONS";
+constexpr char kDetectionBoxesTag[] = "DETECTION_BOXES";
+constexpr char kDetectionListTag[] = "DETECTION_LIST";
+constexpr char kTrackingBoxesTag[] = "TRACKING_BOXES";
+constexpr char kCancelObjectIdTag[] = "CANCEL_OBJECT_ID";
+
+// Move |src| to the back of |dst|.
+void MoveIds(std::vector<int>* dst, std::vector<int> src) {
+  dst->insert(dst->end(), std::make_move_iterator(src.begin()),
+              std::make_move_iterator(src.end()));
+}
+
+int64 GetInputTimestampMs(::mediapipe::CalculatorContext* cc) {
+  return cc->InputTimestamp().Microseconds() / 1000;  // 1 ms = 1000 us.
+}
+
+// Converts a Mediapipe Detection Proto to a TrackedDetection class.
+std::unique_ptr<TrackedDetection> GetTrackedDetectionFromDetection(
+    const Detection& detection, int64 timestamp) {
+  std::unique_ptr<TrackedDetection> tracked_detection =
+      absl::make_unique<TrackedDetection>(detection.detection_id(), timestamp);
+  const float top = detection.location_data().relative_bounding_box().ymin();
+  const float bottom =
+      detection.location_data().relative_bounding_box().ymin() +
+      detection.location_data().relative_bounding_box().height();
+  const float left = detection.location_data().relative_bounding_box().xmin();
+  const float right = detection.location_data().relative_bounding_box().xmin() +
+                      detection.location_data().relative_bounding_box().width();
+  NormalizedRect bounding_box;
+  bounding_box.set_x_center((left + right) / 2.f);
+  bounding_box.set_y_center((top + bottom) / 2.f);
+  bounding_box.set_height(bottom - top);
+  bounding_box.set_width(right - left);
+  tracked_detection->set_bounding_box(bounding_box);
+
+  for (int i = 0; i < detection.label_size(); ++i) {
+    tracked_detection->AddLabel(detection.label(i), detection.score(i));
+  }
+  return tracked_detection;
+}
+
+// Converts a TrackedDetection class to a Mediapipe Detection Proto.
+Detection GetAxisAlignedDetectionFromTrackedDetection(
+    const TrackedDetection& tracked_detection) {
+  Detection detection;
+  LocationData* location_data = detection.mutable_location_data();
+
+  auto corners = tracked_detection.GetCorners();
+
+  float x_min = std::numeric_limits<float>::max();
+  float x_max = std::numeric_limits<float>::min();
+  float y_min = std::numeric_limits<float>::max();
+  float y_max = std::numeric_limits<float>::min();
+  for (int i = 0; i < 4; ++i) {
+    x_min = std::min(x_min, corners[i].x());
+    x_max = std::max(x_max, corners[i].x());
+    y_min = std::min(y_min, corners[i].y());
+    y_max = std::max(y_max, corners[i].y());
+  }
+  location_data->set_format(LocationData::RELATIVE_BOUNDING_BOX);
+  LocationData::RelativeBoundingBox* relative_bbox =
+      location_data->mutable_relative_bounding_box();
+  relative_bbox->set_xmin(x_min);
+  relative_bbox->set_ymin(y_min);
+  relative_bbox->set_width(x_max - x_min);
+  relative_bbox->set_height(y_max - y_min);
+
+  // Use previous id which is the id the object when it's first detected.
+  if (tracked_detection.previous_id() > 0) {
+    detection.set_detection_id(tracked_detection.previous_id());
+  } else {
+    detection.set_detection_id(tracked_detection.unique_id());
+  }
+  for (const auto& label_and_score : tracked_detection.label_to_score_map()) {
+    detection.add_label(label_and_score.first);
+    detection.add_score(label_and_score.second);
+  }
+  return detection;
+}
+
+}  // namespace
+
+// TrackedDetectionManagerCalculator accepts detections and tracking results at
+// different frame rate for real time tracking of targets.
+// Input:
+//   DETECTIONS: A vector<Detection> of newly detected targets.
+//   TRACKING_BOXES: A TimedBoxProtoList which contains a list of tracked boxes
+//   from previous detections.
+//
+// Output:
+//   CANCEL_OBJECT_ID: Ids of targets that are missing/lost such that it should
+//   be removed from tracking.
+//   DETECTIONS: List of detections that are being tracked.
+//   DETECTION_BOXES: List of bounding boxes of detections that are being
+//   tracked.
+//
+// Usage example:
+// node {
+//   calculator: "TrackedDetectionManagerCalculator"
+//   input_stream: "DETECTIONS:detections"
+//   input_stream: "TRACKING_BOXES:boxes"
+//   output_stream: "CANCEL_OBJECT_ID:cancel_object_id"
+//   output_stream: "DETECTIONS:output_detections"
+// }
+class TrackedDetectionManagerCalculator : public CalculatorBase {
+ public:
+  static ::mediapipe::Status GetContract(CalculatorContract* cc);
+
+  ::mediapipe::Status Process(CalculatorContext* cc) override;
+
+ private:
+  // Adds new list of detections to |waiting_for_update_detections_|.
+  void AddDetectionList(const DetectionList& detection_list,
+                        CalculatorContext* cc);
+  void AddDetections(const std::vector<Detection>& detections,
+                     CalculatorContext* cc);
+
+  // Manages existing and new detections.
+  TrackedDetectionManager tracked_detection_manager_;
+
+  // Set of detections that are not up to date yet. These detections will be
+  // added to the detection manager until they got updated from the box tracker.
+  absl::node_hash_map<int, std::unique_ptr<TrackedDetection>>
+      waiting_for_update_detections_;
+};
+REGISTER_CALCULATOR(TrackedDetectionManagerCalculator);
+
+::mediapipe::Status TrackedDetectionManagerCalculator::GetContract(
+    CalculatorContract* cc) {
+  if (cc->Inputs().HasTag(kDetectionsTag)) {
+    cc->Inputs().Tag(kDetectionsTag).Set<std::vector<Detection>>();
+  }
+  if (cc->Inputs().HasTag(kDetectionListTag)) {
+    cc->Inputs().Tag(kDetectionListTag).Set<DetectionList>();
+  }
+  if (cc->Inputs().HasTag(kTrackingBoxesTag)) {
+    cc->Inputs().Tag(kTrackingBoxesTag).Set<TimedBoxProtoList>();
+  }
+
+  if (cc->Outputs().HasTag(kCancelObjectIdTag)) {
+    cc->Outputs().Tag(kCancelObjectIdTag).Set<int>();
+  }
+  if (cc->Outputs().HasTag(kDetectionsTag)) {
+    cc->Outputs().Tag(kDetectionsTag).Set<std::vector<Detection>>();
+  }
+  if (cc->Outputs().HasTag(kDetectionBoxesTag)) {
+    cc->Outputs().Tag(kDetectionBoxesTag).Set<std::vector<NormalizedRect>>();
+  }
+
+  return ::mediapipe::OkStatus();
+}
+
+::mediapipe::Status TrackedDetectionManagerCalculator::Process(
+    CalculatorContext* cc) {
+  if (cc->Inputs().HasTag("TRACKING_BOXES")) {
+    if (!cc->Inputs().Tag("TRACKING_BOXES").IsEmpty()) {
+      const TimedBoxProtoList& tracked_boxes =
+          cc->Inputs().Tag("TRACKING_BOXES").Get<TimedBoxProtoList>();
+
+      // Collect all detections that are removed.
+      auto removed_detection_ids = absl::make_unique<std::vector<int>>();
+      for (const TimedBoxProto& tracked_box : tracked_boxes.box()) {
+        NormalizedRect bounding_box;
+        bounding_box.set_x_center((tracked_box.left() + tracked_box.right()) /
+                                  2.f);
+        bounding_box.set_y_center((tracked_box.bottom() + tracked_box.top()) /
+                                  2.f);
+        bounding_box.set_height(tracked_box.bottom() - tracked_box.top());
+        bounding_box.set_width(tracked_box.right() - tracked_box.left());
+        bounding_box.set_rotation(tracked_box.rotation());
+        // First check if this box updates a detection that's waiting for
+        // update from the tracker.
+        auto waiting_for_update_detectoin_ptr =
+            waiting_for_update_detections_.find(tracked_box.id());
+        if (waiting_for_update_detectoin_ptr !=
+            waiting_for_update_detections_.end()) {
+          // Add the detection and remove duplicated detections.
+          auto removed_ids = tracked_detection_manager_.AddDetection(
+              std::move(waiting_for_update_detectoin_ptr->second));
+          MoveIds(removed_detection_ids.get(), std::move(removed_ids));
+
+          waiting_for_update_detections_.erase(
+              waiting_for_update_detectoin_ptr);
+        }
+        auto removed_ids = tracked_detection_manager_.UpdateDetectionLocation(
+            tracked_box.id(), bounding_box, tracked_box.time_msec());
+        MoveIds(removed_detection_ids.get(), std::move(removed_ids));
+      }
+      // TODO: Should be handled automatically in detection manager.
+      auto removed_ids = tracked_detection_manager_.RemoveObsoleteDetections(
+          GetInputTimestampMs(cc) - kDetectionUpdateTimeOutMS);
+      MoveIds(removed_detection_ids.get(), std::move(removed_ids));
+
+      // TODO: Should be handled automatically in detection manager.
+      removed_ids = tracked_detection_manager_.RemoveOutOfViewDetections();
+      MoveIds(removed_detection_ids.get(), std::move(removed_ids));
+
+      if (!removed_detection_ids->empty() &&
+          cc->Outputs().HasTag(kCancelObjectIdTag)) {
+        auto timestamp = cc->InputTimestamp();
+        for (int box_id : *removed_detection_ids) {
+          // The timestamp is incremented (by 1 us) because currently the box
+          // tracker calculator only accepts one cancel object ID for any given
+          // timestamp.
+          cc->Outputs()
+              .Tag(kCancelObjectIdTag)
+              .AddPacket(mediapipe::MakePacket<int>(box_id).At(timestamp++));
+        }
+      }
+
+      // Output detections and corresponding bounding boxes.
+      const auto& all_detections =
+          tracked_detection_manager_.GetAllTrackedDetections();
+      auto output_detections = absl::make_unique<std::vector<Detection>>();
+      auto output_boxes = absl::make_unique<std::vector<NormalizedRect>>();
+
+      for (const auto& detection_ptr : all_detections) {
+        const auto& detection = *detection_ptr.second;
+        // Only output detections that are synced.
+        if (detection.last_updated_timestamp() <
+            cc->InputTimestamp().Microseconds() / 1000) {
+          continue;
+        }
+        output_detections->emplace_back(
+            GetAxisAlignedDetectionFromTrackedDetection(detection));
+        output_boxes->emplace_back(detection.bounding_box());
+      }
+      if (cc->Outputs().HasTag(kDetectionsTag)) {
+        cc->Outputs()
+            .Tag(kDetectionsTag)
+            .Add(output_detections.release(), cc->InputTimestamp());
+      }
+
+      if (cc->Outputs().HasTag(kDetectionBoxesTag)) {
+        cc->Outputs()
+            .Tag(kDetectionBoxesTag)
+            .Add(output_boxes.release(), cc->InputTimestamp());
+      }
+    }
+  }
+
+  if (cc->Inputs().HasTag(kDetectionsTag) &&
+      !cc->Inputs().Tag(kDetectionsTag).IsEmpty()) {
+    const auto detections =
+        cc->Inputs().Tag(kDetectionsTag).Get<std::vector<Detection>>();
+    AddDetections(detections, cc);
+  }
+
+  if (cc->Inputs().HasTag(kDetectionListTag) &&
+      !cc->Inputs().Tag(kDetectionListTag).IsEmpty()) {
+    const auto detection_list =
+        cc->Inputs().Tag(kDetectionListTag).Get<DetectionList>();
+    AddDetectionList(detection_list, cc);
+  }
+
+  return ::mediapipe::OkStatus();
+}
+
+void TrackedDetectionManagerCalculator::AddDetectionList(
+    const DetectionList& detection_list, CalculatorContext* cc) {
+  for (const auto& detection : detection_list.detection()) {
+    // Convert from microseconds to milliseconds.
+    std::unique_ptr<TrackedDetection> new_detection =
+        GetTrackedDetectionFromDetection(
+            detection, cc->InputTimestamp().Microseconds() / 1000);
+
+    const int id = new_detection->unique_id();
+    waiting_for_update_detections_[id] = std::move(new_detection);
+  }
+}
+
+void TrackedDetectionManagerCalculator::AddDetections(
+    const std::vector<Detection>& detections, CalculatorContext* cc) {
+  for (const auto& detection : detections) {
+    // Convert from microseconds to milliseconds.
+    std::unique_ptr<TrackedDetection> new_detection =
+        GetTrackedDetectionFromDetection(
+            detection, cc->InputTimestamp().Microseconds() / 1000);
+
+    const int id = new_detection->unique_id();
+    waiting_for_update_detections_[id] = std::move(new_detection);
+  }
+}
+
+}  // namespace mediapipe

mediapipe/calculators/video/tracking_graph_test.cc

@@ -0,0 +1,709 @@
+// 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 <fstream>
+#include <map>
+#include <memory>
+#include <string>
+#include <utility>
+#include <vector>
+
+#include "mediapipe/calculators/video/box_tracker_calculator.pb.h"
+#include "mediapipe/framework/calculator.pb.h"
+#include "mediapipe/framework/calculator_framework.h"
+#include "mediapipe/framework/deps/file_path.h"
+#include "mediapipe/framework/formats/image_frame.h"
+#include "mediapipe/framework/packet.h"
+#include "mediapipe/framework/port/advanced_proto_inc.h"
+#include "mediapipe/framework/port/file_helpers.h"
+#include "mediapipe/framework/port/gmock.h"
+#include "mediapipe/framework/port/gtest.h"
+#include "mediapipe/framework/port/opencv_highgui_inc.h"
+#include "mediapipe/framework/port/proto_ns.h"
+#include "mediapipe/framework/port/status.h"
+#include "mediapipe/framework/port/status_matchers.h"
+#include "mediapipe/util/tracking/box_tracker.pb.h"
+#include "mediapipe/util/tracking/tracking.pb.h"
+
+#ifdef __APPLE__
+#include <CoreFoundation/CoreFoundation.h>
+#endif  // defined(__APPLE__)
+
+namespace mediapipe {
+namespace {
+using ::testing::FloatNear;
+using ::testing::Test;
+
+std::string GetTestDir() {
+#ifdef __APPLE__
+  char path[1024];
+  CFURLRef bundle_url = CFBundleCopyBundleURL(CFBundleGetMainBundle());
+  CFURLGetFileSystemRepresentation(
+      bundle_url, true, reinterpret_cast<UInt8*>(path), sizeof(path));
+  CFRelease(bundle_url);
+  return ::mediapipe::file::JoinPath(path, "testdata");
+#elif defined(__ANDROID__)
+  char path[1024];
+  getcwd(path, sizeof(path));
+  return ::mediapipe::file::JoinPath(path,
+                                     "mediapipe/calculators/video/testdata");
+#else
+  return ::mediapipe::file::JoinPath(
+      "./",
+      // This should match the path of the output files
+      // of the genrule() that generates test model files.
+      "mediapipe/calculators/video/testdata");
+#endif  // defined(__APPLE__)
+}
+
+bool LoadBinaryTestGraph(const std::string& graph_path,
+                         CalculatorGraphConfig* config) {
+  std::ifstream ifs;
+  ifs.open(graph_path.c_str());
+  proto_ns::io::IstreamInputStream in_stream(&ifs);
+  bool success = config->ParseFromZeroCopyStream(&in_stream);
+  ifs.close();
+  if (!success) {
+    LOG(ERROR) << "could not parse test graph: " << graph_path;
+  }
+  return success;
+}
+
+class TrackingGraphTest : public Test {
+ protected:
+  TrackingGraphTest() {}
+
+  void SetUp() override {
+    test_dir_ = GetTestDir();
+    const auto graph_path = file::JoinPath(test_dir_, "tracker.binarypb");
+    ASSERT_TRUE(LoadBinaryTestGraph(graph_path, &config_));
+
+    original_image_ = cv::imread(file::JoinPath(test_dir_, "lenna.png"));
+    CreateInputFramesFromOriginalImage(kNumImages, kTranslationStep,
+                                       &input_frames_packets_);
+
+    const auto& first_input_img = input_frames_packets_[0].Get<ImageFrame>();
+    const int img_width = first_input_img.Width();
+    const int img_height = first_input_img.Height();
+    translation_step_x_ = kTranslationStep / static_cast<float>(img_width);
+    translation_step_y_ = kTranslationStep / static_cast<float>(img_height);
+
+    // Creat new configure and packet dump vector to store output.
+    mediapipe::CalculatorGraphConfig config_copy = config_;
+    mediapipe::tool::AddVectorSink("boxes", &config_copy, &output_packets_);
+    mediapipe::tool::AddVectorSink("ra_boxes", &config_copy,
+                                   &random_access_results_packets_);
+
+    // Initialize graph
+    MP_ASSERT_OK(graph_.Initialize(config_copy));
+
+    const auto parallel_graph_path =
+        file::JoinPath(test_dir_, "parallel_tracker.binarypb");
+    CalculatorGraphConfig parallel_config;
+    ASSERT_TRUE(LoadBinaryTestGraph(parallel_graph_path, &parallel_config));
+    mediapipe::tool::AddVectorSink("boxes", &parallel_config, &output_packets_);
+    mediapipe::tool::AddVectorSink("ra_boxes", &parallel_config,
+                                   &random_access_results_packets_);
+    MP_ASSERT_OK(parallel_graph_.Initialize(parallel_config));
+  }
+
+  void CreateInputFramesFromOriginalImage(
+      int num_images, int translation_step,
+      std::vector<Packet>* input_frames_packets);
+
+  void TearDown() override {
+    output_packets_.clear();
+    random_access_results_packets_.clear();
+  }
+
+  std::unique_ptr<TimedBoxProtoList> MakeBoxList(
+      const Timestamp& timestamp, const std::vector<bool>& is_quad_tracking,
+      const std::vector<bool>& is_pnp_tracking,
+      const std::vector<bool>& reacquisition) const;
+
+  void RunGraphWithSidePacketsAndInputs(
+      const std::map<std::string, mediapipe::Packet>& side_packets,
+      const mediapipe::Packet& start_pos_packet);
+
+  // Utility functions used to judge if a given quad or box is near to the
+  // groundtruth location at a given frame.
+  // Examine box.reacquisition() field equals to `reacquisition`.
+  // `frame` can be float number to account for inter-frame interpolation.
+  void ExpectBoxAtFrame(const TimedBoxProto& box, float frame,
+                        bool reacquisition);
+
+  // Examine box.aspect_ratio() field equals to `aspect_ratio` if asepct_ratio
+  // is positive.
+  void ExpectQuadAtFrame(const TimedBoxProto& box, float frame,
+                         float aspect_ratio, bool reacquisition);
+
+  // Utility function to judge if two quad are near to each other.
+  void ExpectQuadNear(const TimedBoxProto& box1, const TimedBoxProto& box2);
+
+  std::unique_ptr<TimedBoxProtoList> CreateRandomAccessTrackingBoxList(
+      const std::vector<Timestamp>& start_timestamps,
+      const std::vector<Timestamp>& end_timestamps) const;
+
+  CalculatorGraph graph_;
+  CalculatorGraph parallel_graph_;
+  CalculatorGraphConfig config_;
+  std::string test_dir_;
+  cv::Mat original_image_;
+  std::vector<Packet> input_frames_packets_;
+  std::vector<mediapipe::Packet> output_packets_;
+  std::vector<mediapipe::Packet> random_access_results_packets_;
+  float translation_step_x_;  // normalized translation step in x direction
+  float translation_step_y_;  // normalized translation step in y direction
+  static constexpr float kInitialBoxHalfWidthNormalized = 0.25f;
+  static constexpr float kInitialBoxHalfHeightNormalized = 0.25f;
+  static constexpr float kImageAspectRatio = 1.0f;  // for lenna.png
+  static constexpr float kInitialBoxLeft =
+      0.5f - kInitialBoxHalfWidthNormalized;
+  static constexpr float kInitialBoxRight =
+      0.5f + kInitialBoxHalfWidthNormalized;
+  static constexpr float kInitialBoxTop =
+      0.5f - kInitialBoxHalfHeightNormalized;
+  static constexpr float kInitialBoxBottom =
+      0.5f + kInitialBoxHalfHeightNormalized;
+  static constexpr int kFrameIntervalUs = 30000;
+  static constexpr int kNumImages = 8;
+  // Each image is shifted to the right and bottom by kTranslationStep
+  // pixels compared with the previous image.
+  static constexpr int kTranslationStep = 10;
+  static constexpr float kEqualityTolerance = 3e-4f;
+};
+
+void TrackingGraphTest::ExpectBoxAtFrame(const TimedBoxProto& box, float frame,
+                                         bool reacquisition) {
+  EXPECT_EQ(box.reacquisition(), reacquisition);
+  EXPECT_TRUE(box.has_rotation());
+  EXPECT_THAT(box.rotation(), FloatNear(0, kEqualityTolerance));
+  EXPECT_THAT(box.left(),
+              FloatNear(kInitialBoxLeft - frame * translation_step_x_,
+                        kEqualityTolerance));
+  EXPECT_THAT(box.top(), FloatNear(kInitialBoxTop - frame * translation_step_y_,
+                                   kEqualityTolerance));
+  EXPECT_THAT(box.bottom(),
+              FloatNear(kInitialBoxBottom - frame * translation_step_y_,
+                        kEqualityTolerance));
+  EXPECT_THAT(box.right(),
+              FloatNear(kInitialBoxRight - frame * translation_step_x_,
+                        kEqualityTolerance));
+}
+
+void TrackingGraphTest::ExpectQuadAtFrame(const TimedBoxProto& box, float frame,
+                                          float aspect_ratio,
+                                          bool reacquisition) {
+  EXPECT_TRUE(box.has_quad()) << "quad must exist!";
+  if (aspect_ratio > 0) {
+    EXPECT_TRUE(box.has_aspect_ratio());
+    EXPECT_NEAR(box.aspect_ratio(), aspect_ratio, kEqualityTolerance);
+  }
+
+  EXPECT_EQ(box.reacquisition(), reacquisition);
+
+  const auto& quad = box.quad();
+  EXPECT_EQ(8, quad.vertices_size())
+      << "quad has only " << box.quad().vertices_size() << " vertices";
+  EXPECT_THAT(quad.vertices(0),
+              FloatNear(kInitialBoxLeft - frame * translation_step_x_,
+                        kEqualityTolerance));
+  EXPECT_THAT(quad.vertices(1),
+              FloatNear(kInitialBoxTop - frame * translation_step_y_,
+                        kEqualityTolerance));
+  EXPECT_THAT(quad.vertices(3),
+              FloatNear(kInitialBoxBottom - frame * translation_step_y_,
+                        kEqualityTolerance));
+  EXPECT_THAT(quad.vertices(4),
+              FloatNear(kInitialBoxRight - frame * translation_step_x_,
+                        kEqualityTolerance));
+}
+
+void TrackingGraphTest::ExpectQuadNear(const TimedBoxProto& box1,
+                                       const TimedBoxProto& box2) {
+  EXPECT_TRUE(box1.has_quad());
+  EXPECT_TRUE(box2.has_quad());
+  EXPECT_EQ(8, box1.quad().vertices_size())
+      << "quad has only " << box1.quad().vertices_size() << " vertices";
+  EXPECT_EQ(8, box2.quad().vertices_size())
+      << "quad has only " << box2.quad().vertices_size() << " vertices";
+  for (int j = 0; j < box1.quad().vertices_size(); ++j) {
+    EXPECT_NEAR(box1.quad().vertices(j), box2.quad().vertices(j),
+                kEqualityTolerance);
+  }
+}
+
+std::unique_ptr<TimedBoxProtoList> TrackingGraphTest::MakeBoxList(
+    const Timestamp& timestamp, const std::vector<bool>& is_quad_tracking,
+    const std::vector<bool>& is_pnp_tracking,
+    const std::vector<bool>& reacquisition) const {
+  auto box_list = absl::make_unique<TimedBoxProtoList>();
+  int box_id = 0;
+  for (int j = 0; j < is_quad_tracking.size(); ++j) {
+    TimedBoxProto* box = box_list->add_box();
+    if (is_quad_tracking[j]) {
+      box->mutable_quad()->add_vertices(kInitialBoxLeft);
+      box->mutable_quad()->add_vertices(kInitialBoxTop);
+      box->mutable_quad()->add_vertices(kInitialBoxLeft);
+      box->mutable_quad()->add_vertices(kInitialBoxBottom);
+      box->mutable_quad()->add_vertices(kInitialBoxRight);
+      box->mutable_quad()->add_vertices(kInitialBoxBottom);
+      box->mutable_quad()->add_vertices(kInitialBoxRight);
+      box->mutable_quad()->add_vertices(kInitialBoxTop);
+
+      if (is_pnp_tracking[j]) {
+        box->set_aspect_ratio(kImageAspectRatio);
+      }
+    } else {
+      box->set_left(kInitialBoxLeft);
+      box->set_right(kInitialBoxRight);
+      box->set_top(kInitialBoxTop);
+      box->set_bottom(kInitialBoxBottom);
+    }
+    box->set_id(box_id++);
+    box->set_time_msec(timestamp.Value() / 1000);
+    box->set_reacquisition(reacquisition[j]);
+  }
+
+  return box_list;
+}
+
+void TrackingGraphTest::CreateInputFramesFromOriginalImage(
+    int num_images, int translation_step,
+    std::vector<Packet>* input_frames_packets) {
+  const int crop_width = original_image_.cols - num_images * translation_step;
+  const int crop_height = original_image_.rows - num_images * translation_step;
+  for (int i = 0; i < num_images; ++i) {
+    cv::Rect roi(i * translation_step, i * translation_step, crop_width,
+                 crop_height);
+    cv::Mat cropped_img = cv::Mat(original_image_, roi);
+    auto cropped_image_frame = absl::make_unique<ImageFrame>(
+        ImageFormat::SRGB, crop_width, crop_height, cropped_img.step[0],
+        cropped_img.data, ImageFrame::PixelDataDeleter::kNone);
+    Timestamp curr_timestamp = Timestamp(i * kFrameIntervalUs);
+    Packet image_packet =
+        Adopt(cropped_image_frame.release()).At(curr_timestamp);
+    input_frames_packets->push_back(image_packet);
+  }
+}
+
+void TrackingGraphTest::RunGraphWithSidePacketsAndInputs(
+    const std::map<std::string, mediapipe::Packet>& side_packets,
+    const mediapipe::Packet& start_pos_packet) {
+  // Start running the graph
+  MP_EXPECT_OK(graph_.StartRun(side_packets));
+
+  MP_EXPECT_OK(graph_.AddPacketToInputStream("start_pos", start_pos_packet));
+
+  for (auto frame_packet : input_frames_packets_) {
+    MP_EXPECT_OK(
+        graph_.AddPacketToInputStream("image_cpu_frames", frame_packet));
+    MP_EXPECT_OK(graph_.WaitUntilIdle());
+  }
+
+  MP_EXPECT_OK(graph_.CloseAllInputStreams());
+  MP_EXPECT_OK(graph_.WaitUntilDone());
+}
+
+std::unique_ptr<TimedBoxProtoList>
+TrackingGraphTest::CreateRandomAccessTrackingBoxList(
+    const std::vector<Timestamp>& start_timestamps,
+    const std::vector<Timestamp>& end_timestamps) const {
+  CHECK_EQ(start_timestamps.size(), end_timestamps.size());
+  auto ra_boxes = absl::make_unique<TimedBoxProtoList>();
+  for (int i = 0; i < start_timestamps.size(); ++i) {
+    auto start_box_list =
+        MakeBoxList(start_timestamps[i], std::vector<bool>{true},
+                    std::vector<bool>{true}, std::vector<bool>{false});
+    auto end_box_list =
+        MakeBoxList(end_timestamps[i], std::vector<bool>{true},
+                    std::vector<bool>{true}, std::vector<bool>{false});
+    *(ra_boxes->add_box()) = (*start_box_list).box(0);
+    *(ra_boxes->add_box()) = (*end_box_list).box(0);
+  }
+  return ra_boxes;
+}
+
+TEST_F(TrackingGraphTest, BasicBoxTrackingSanityCheck) {
+  // Create input side packets.
+  std::map<std::string, mediapipe::Packet> side_packets;
+  side_packets.insert(std::make_pair("analysis_downsample_factor",
+                                     mediapipe::MakePacket<float>(1.0f)));
+  side_packets.insert(std::make_pair(
+      "calculator_options",
+      mediapipe::MakePacket<CalculatorOptions>(CalculatorOptions())));
+
+  // Run the graph with input side packets, start_pos, and input image frames.
+  Timestamp start_box_time = input_frames_packets_[0].Timestamp();
+  // is_quad_tracking is used to indicate whether to track quad for each
+  // individual box.
+  std::vector<bool> is_quad_tracking{false};
+  // is_pnp_tracking is used to indicate whether to use perspective transform to
+  // track quad.
+  std::vector<bool> is_pnp_tracking{false};
+  // is_reacquisition is used to indicate whether to enable reacquisition for
+  // the box.
+  std::vector<bool> is_reacquisition{false};
+  auto start_box_list = MakeBoxList(start_box_time, is_quad_tracking,
+                                    is_pnp_tracking, is_reacquisition);
+  Packet start_pos_packet = Adopt(start_box_list.release()).At(start_box_time);
+  RunGraphWithSidePacketsAndInputs(side_packets, start_pos_packet);
+
+  EXPECT_EQ(input_frames_packets_.size(), output_packets_.size());
+
+  for (int i = 0; i < output_packets_.size(); ++i) {
+    const TimedBoxProtoList& boxes =
+        output_packets_[i].Get<TimedBoxProtoList>();
+    EXPECT_EQ(is_quad_tracking.size(), boxes.box_size());
+    ExpectBoxAtFrame(boxes.box(0), i, false);
+  }
+}
+
+TEST_F(TrackingGraphTest, BasicQuadTrackingSanityCheck) {
+  // Create input side packets.
+  std::map<std::string, mediapipe::Packet> side_packets;
+  side_packets.insert(std::make_pair("analysis_downsample_factor",
+                                     mediapipe::MakePacket<float>(1.0f)));
+  CalculatorOptions calculator_options;
+  calculator_options.MutableExtension(BoxTrackerCalculatorOptions::ext)
+      ->mutable_tracker_options()
+      ->mutable_track_step_options()
+      ->set_tracking_degrees(
+          TrackStepOptions::TRACKING_DEGREE_OBJECT_PERSPECTIVE);
+  side_packets.insert(std::make_pair(
+      "calculator_options",
+      mediapipe::MakePacket<CalculatorOptions>(calculator_options)));
+
+  Timestamp start_box_time = input_frames_packets_[0].Timestamp();
+  // Box id 0 use quad tracking with 8DoF homography transform.
+  // Box id 1 use quad tracking with 6DoF perspective transform.
+  // Box id 2 use box tracking with 4DoF similarity transform.
+  std::vector<bool> is_quad_tracking{true, true, false};
+  std::vector<bool> is_pnp_tracking{false, true, false};
+  std::vector<bool> is_reacquisition{true, false, true};
+  auto start_box_list = MakeBoxList(start_box_time, is_quad_tracking,
+                                    is_pnp_tracking, is_reacquisition);
+  Packet start_pos_packet = Adopt(start_box_list.release()).At(start_box_time);
+  RunGraphWithSidePacketsAndInputs(side_packets, start_pos_packet);
+
+  EXPECT_EQ(input_frames_packets_.size(), output_packets_.size());
+  for (int i = 0; i < output_packets_.size(); ++i) {
+    const TimedBoxProtoList& boxes =
+        output_packets_[i].Get<TimedBoxProtoList>();
+    EXPECT_EQ(is_quad_tracking.size(), boxes.box_size());
+    for (int j = 0; j < boxes.box_size(); ++j) {
+      const TimedBoxProto& box = boxes.box(j);
+      if (is_quad_tracking[box.id()]) {
+        ExpectQuadAtFrame(box, i,
+                          is_pnp_tracking[box.id()] ? kImageAspectRatio : -1.0f,
+                          is_reacquisition[box.id()]);
+      } else {
+        ExpectBoxAtFrame(box, i, is_reacquisition[box.id()]);
+      }
+    }
+  }
+}
+
+TEST_F(TrackingGraphTest, TestRandomAccessTrackingResults) {
+  // Create input side packets.
+  std::map<std::string, mediapipe::Packet> side_packets;
+  side_packets.insert(std::make_pair("analysis_downsample_factor",
+                                     mediapipe::MakePacket<float>(1.0f)));
+  CalculatorOptions calculator_options;
+  calculator_options.MutableExtension(BoxTrackerCalculatorOptions::ext)
+      ->mutable_tracker_options()
+      ->mutable_track_step_options()
+      ->set_tracking_degrees(
+          TrackStepOptions::TRACKING_DEGREE_OBJECT_PERSPECTIVE);
+  side_packets.insert(std::make_pair(
+      "calculator_options",
+      mediapipe::MakePacket<CalculatorOptions>(calculator_options)));
+
+  ASSERT_GT(input_frames_packets_.size(), 2);  // at least 3 frames
+  ASSERT_TRUE(input_frames_packets_[2].Timestamp() -
+                  input_frames_packets_[1].Timestamp() >
+              TimestampDiff(1000));
+
+  constexpr int start_frame = 0;
+  Timestamp start_box_time = input_frames_packets_[start_frame].Timestamp();
+  auto start_box_list =
+      MakeBoxList(start_box_time, std::vector<bool>{true},
+                  std::vector<bool>{true}, std::vector<bool>{false});
+  constexpr int end_frame = 2;
+  Timestamp end_box_time = input_frames_packets_[end_frame].Timestamp();
+
+  // Also test reverse random access tracking.
+  // This offset of 1ms is simulating the case where the start query timestamp
+  // to be not any existing frame timestamp. In reality, it's highly encouraged
+  // to have the start query timestamp be aligned with frame timestamp.
+  constexpr int reverse_start_frame = 1;
+  Timestamp reverse_start_box_time =
+      input_frames_packets_[reverse_start_frame].Timestamp() + 1000;
+
+  auto ra_boxes = CreateRandomAccessTrackingBoxList(
+      {start_box_time, reverse_start_box_time}, {end_box_time, start_box_time});
+
+  Packet ra_packet = Adopt(ra_boxes.release()).At(start_box_time);
+  Packet start_packet = Adopt(start_box_list.release()).At(start_box_time);
+
+  // Start running the ordinary graph, verify random access produce same result
+  // as normal tracking.
+  MP_EXPECT_OK(graph_.StartRun(side_packets));
+  MP_EXPECT_OK(graph_.AddPacketToInputStream("start_pos", start_packet));
+  for (auto frame_packet : input_frames_packets_) {
+    MP_EXPECT_OK(
+        graph_.AddPacketToInputStream("image_cpu_frames", frame_packet));
+    Packet track_time_packet = Adopt(new int(0)).At(frame_packet.Timestamp());
+    MP_EXPECT_OK(
+        graph_.AddPacketToInputStream("track_time", track_time_packet));
+    MP_EXPECT_OK(graph_.WaitUntilIdle());
+  }
+  MP_EXPECT_OK(graph_.AddPacketToInputStream("ra_track", ra_packet));
+  MP_EXPECT_OK(graph_.CloseAllInputStreams());
+  MP_EXPECT_OK(graph_.WaitUntilDone());
+
+  EXPECT_EQ(input_frames_packets_.size(), output_packets_.size());
+  const TimedBoxProtoList tracking_result =
+      output_packets_[end_frame].Get<TimedBoxProtoList>();
+  EXPECT_EQ(1, tracking_result.box_size());
+
+  // Should have 1 random access packet.
+  EXPECT_EQ(1, random_access_results_packets_.size());
+  const TimedBoxProtoList& ra_result =
+      random_access_results_packets_[0].Get<TimedBoxProtoList>();
+  // Two box tracking results. One for comparison with normal tracking. The
+  // other for reverse random access tracking.
+  EXPECT_EQ(2, ra_result.box_size());
+
+  // Check if randan access tracking has same result with normal tracking.
+  ExpectQuadNear(tracking_result.box(0), ra_result.box(0));
+  ExpectQuadAtFrame(ra_result.box(0), end_frame - start_frame,
+                    kImageAspectRatio, false);
+  ExpectQuadAtFrame(ra_result.box(1), start_frame - reverse_start_frame - 1,
+                    kImageAspectRatio, false);
+
+  // Clear output and ra result packet vector before test parallel graph.
+  TearDown();
+
+  // Start running the parallel graph, verify random access produce same result
+  // as normal tracking.
+  MP_EXPECT_OK(parallel_graph_.StartRun(side_packets));
+  MP_EXPECT_OK(
+      parallel_graph_.AddPacketToInputStream("start_pos", start_packet));
+  for (auto frame_packet : input_frames_packets_) {
+    MP_EXPECT_OK(parallel_graph_.AddPacketToInputStream("image_cpu_frames",
+                                                        frame_packet));
+    MP_EXPECT_OK(parallel_graph_.WaitUntilIdle());
+  }
+  MP_EXPECT_OK(parallel_graph_.AddPacketToInputStream("ra_track", ra_packet));
+  MP_EXPECT_OK(parallel_graph_.CloseAllInputStreams());
+  MP_EXPECT_OK(parallel_graph_.WaitUntilDone());
+
+  EXPECT_EQ(input_frames_packets_.size(), output_packets_.size());
+  const TimedBoxProtoList parallel_tracking_result =
+      output_packets_[end_frame].Get<TimedBoxProtoList>();
+  EXPECT_EQ(1, parallel_tracking_result.box_size());
+
+  // should have only 1 random access
+  EXPECT_EQ(1, random_access_results_packets_.size());
+  const TimedBoxProtoList& parallel_ra_result =
+      random_access_results_packets_[0].Get<TimedBoxProtoList>();
+  EXPECT_EQ(2, parallel_ra_result.box_size());
+
+  // Check if randan access tracking has same result with normal tracking.
+  ExpectQuadNear(parallel_tracking_result.box(0), parallel_ra_result.box(0));
+  ExpectQuadAtFrame(parallel_ra_result.box(0), end_frame - start_frame,
+                    kImageAspectRatio, false);
+  ExpectQuadAtFrame(parallel_ra_result.box(1),
+                    start_frame - reverse_start_frame - 1, kImageAspectRatio,
+                    false);
+}
+
+// Tests what happens when random access request timestamps are
+// outside of cache.
+TEST_F(TrackingGraphTest, TestRandomAccessTrackingTimestamps) {
+  // Create input side packets.
+  std::map<std::string, mediapipe::Packet> side_packets;
+  side_packets.insert(std::make_pair("analysis_downsample_factor",
+                                     mediapipe::MakePacket<float>(1.0f)));
+  CalculatorOptions calculator_options;
+  calculator_options.MutableExtension(BoxTrackerCalculatorOptions::ext)
+      ->mutable_tracker_options()
+      ->mutable_track_step_options()
+      ->set_tracking_degrees(
+          TrackStepOptions::TRACKING_DEGREE_OBJECT_PERSPECTIVE);
+  // We intentionally don't cache all frames, to see what happens when
+  // random access tracking request time falls outside cache range.
+  calculator_options.MutableExtension(BoxTrackerCalculatorOptions::ext)
+      ->set_streaming_track_data_cache_size(input_frames_packets_.size() - 1);
+  side_packets.insert(std::make_pair(
+      "calculator_options",
+      mediapipe::MakePacket<CalculatorOptions>(calculator_options)));
+
+  // Set up random access boxes
+  const int num_frames = input_frames_packets_.size();
+  const int64 usec_in_sec = 1000000;
+  std::vector<Timestamp> start_timestamps{
+      input_frames_packets_[0].Timestamp() - usec_in_sec,  // forward
+      input_frames_packets_[0].Timestamp(),                // forward
+      input_frames_packets_[1].Timestamp(),                // forward
+      input_frames_packets_[num_frames - 1].Timestamp() + usec_in_sec,  // fwd
+      input_frames_packets_[0].Timestamp(),               // backward
+      input_frames_packets_[num_frames - 1].Timestamp(),  // backward
+      input_frames_packets_[num_frames - 1].Timestamp(),  // backward
+      input_frames_packets_[num_frames - 1].Timestamp() + usec_in_sec  // back
+  };
+  std::vector<Timestamp> end_timestamps{
+      input_frames_packets_[num_frames - 1].Timestamp(),
+      input_frames_packets_[num_frames - 1].Timestamp(),
+      input_frames_packets_[num_frames - 1].Timestamp() + usec_in_sec,
+      input_frames_packets_[num_frames - 1].Timestamp() + 2 * usec_in_sec,
+      input_frames_packets_[0].Timestamp() - usec_in_sec,
+      input_frames_packets_[0].Timestamp(),
+      input_frames_packets_[0].Timestamp() - usec_in_sec,
+      input_frames_packets_[1].Timestamp()};
+  auto ra_boxes =
+      CreateRandomAccessTrackingBoxList(start_timestamps, end_timestamps);
+  Packet ra_packet =
+      Adopt(ra_boxes.release()).At(input_frames_packets_[0].Timestamp());
+
+  // Run the graph and check if the outside-cache request have no results.
+  // Start running the parallel graph, verify random access produce same result
+  // as normal tracking.
+  MP_EXPECT_OK(parallel_graph_.StartRun(side_packets));
+  for (auto frame_packet : input_frames_packets_) {
+    MP_EXPECT_OK(parallel_graph_.AddPacketToInputStream("image_cpu_frames",
+                                                        frame_packet));
+    MP_EXPECT_OK(parallel_graph_.WaitUntilIdle());
+  }
+  MP_EXPECT_OK(parallel_graph_.AddPacketToInputStream("ra_track", ra_packet));
+  MP_EXPECT_OK(parallel_graph_.CloseAllInputStreams());
+  MP_EXPECT_OK(parallel_graph_.WaitUntilDone());
+
+  // should have 1 random access packet with 0 result boxes
+  EXPECT_EQ(1, random_access_results_packets_.size());
+  const auto& ra_returned_boxes =
+      random_access_results_packets_[0].Get<TimedBoxProtoList>();
+  const int num_returned_ra_boxes = ra_returned_boxes.box_size();
+  EXPECT_EQ(0, num_returned_ra_boxes);
+}
+
+TEST_F(TrackingGraphTest, TestTransitionFramesForReacquisition) {
+  // Create input side packets.
+  std::map<std::string, mediapipe::Packet> side_packets;
+  side_packets.insert(std::make_pair("analysis_downsample_factor",
+                                     mediapipe::MakePacket<float>(1.0f)));
+  CalculatorOptions calculator_options;
+  calculator_options.MutableExtension(BoxTrackerCalculatorOptions::ext)
+      ->mutable_tracker_options()
+      ->mutable_track_step_options()
+      ->set_tracking_degrees(
+          TrackStepOptions::TRACKING_DEGREE_OBJECT_PERSPECTIVE);
+  constexpr int kTransitionFrames = 3;
+  calculator_options.MutableExtension(BoxTrackerCalculatorOptions::ext)
+      ->set_start_pos_transition_frames(kTransitionFrames);
+
+  side_packets.insert(std::make_pair(
+      "calculator_options",
+      mediapipe::MakePacket<CalculatorOptions>(calculator_options)));
+
+  Timestamp start_box_time = input_frames_packets_[0].Timestamp();
+  // Box id 0 use quad tracking with 8DoF homography transform.
+  // Box id 1 use quad tracking with 6DoF perspective transform.
+  // Box id 2 use box tracking with 4DoF similarity transform.
+  std::vector<bool> is_quad_tracking{true, true, false};
+  std::vector<bool> is_pnp_tracking{false, true, false};
+  std::vector<bool> is_reacquisition{true, true, true};
+  auto start_box_list = MakeBoxList(start_box_time, is_quad_tracking,
+                                    is_pnp_tracking, is_reacquisition);
+  Packet start_pos_packet = Adopt(start_box_list.release()).At(start_box_time);
+
+  // Setting box pos restart from initial position (frame 0's position).
+  constexpr int kRestartFrame = 3;
+  Timestamp restart_box_time = input_frames_packets_[kRestartFrame].Timestamp();
+  auto restart_box_list = MakeBoxList(restart_box_time, is_quad_tracking,
+                                      is_pnp_tracking, is_reacquisition);
+  Packet restart_pos_packet =
+      Adopt(restart_box_list.release()).At(restart_box_time);
+  MP_EXPECT_OK(graph_.StartRun(side_packets));
+  MP_EXPECT_OK(graph_.AddPacketToInputStream("start_pos", start_pos_packet));
+
+  for (int j = 0; j < input_frames_packets_.size(); ++j) {
+    // Add TRACK_TIME stream queries in between 2 frames.
+    if (j > 0) {
+      Timestamp track_time = Timestamp((j - 0.5f) * kFrameIntervalUs);
+      LOG(INFO) << track_time.Value();
+      Packet track_time_packet = Adopt(new Timestamp).At(track_time);
+      MP_EXPECT_OK(
+          graph_.AddPacketToInputStream("track_time", track_time_packet));
+    }
+
+    MP_EXPECT_OK(graph_.AddPacketToInputStream("image_cpu_frames",
+                                               input_frames_packets_[j]));
+    Packet track_time_packet =
+        Adopt(new int(0)).At(input_frames_packets_[j].Timestamp());
+    MP_EXPECT_OK(
+        graph_.AddPacketToInputStream("track_time", track_time_packet));
+    MP_EXPECT_OK(graph_.WaitUntilIdle());
+
+    if (j == kRestartFrame) {
+      MP_EXPECT_OK(
+          graph_.AddPacketToInputStream("restart_pos", restart_pos_packet));
+    }
+  }
+
+  MP_EXPECT_OK(graph_.CloseAllInputStreams());
+  MP_EXPECT_OK(graph_.WaitUntilDone());
+
+  EXPECT_EQ(input_frames_packets_.size() * 2 - 1, output_packets_.size());
+  for (int i = 0; i < output_packets_.size(); ++i) {
+    const TimedBoxProtoList& boxes =
+        output_packets_[i].Get<TimedBoxProtoList>();
+    EXPECT_EQ(is_quad_tracking.size(), boxes.box_size());
+    float frame_id = i / 2.0f;
+    float expected_frame_id;
+    if (frame_id <= kRestartFrame) {
+      // before transition
+      expected_frame_id = frame_id;
+    } else {
+      float transition_frames = frame_id - kRestartFrame;
+      if (transition_frames <= kTransitionFrames) {
+        // transitioning.
+        expected_frame_id =
+            kRestartFrame -
+            transition_frames / kTransitionFrames * kRestartFrame +
+            transition_frames;
+      } else {
+        // after transition.
+        expected_frame_id = transition_frames;
+      }
+    }
+
+    for (int j = 0; j < boxes.box_size(); ++j) {
+      const TimedBoxProto& box = boxes.box(j);
+      if (is_quad_tracking[box.id()]) {
+        ExpectQuadAtFrame(box, expected_frame_id,
+                          is_pnp_tracking[box.id()] ? kImageAspectRatio : -1.0f,
+                          is_reacquisition[box.id()]);
+      } else {
+        ExpectBoxAtFrame(box, expected_frame_id, is_reacquisition[box.id()]);
+      }
+    }
+  }
+}
+
+// TODO: Add test for reacquisition.
+
+}  // namespace
+}  // namespace mediapipe