Added more benchmark scripts for the Tasks Python API

mediapipe/tasks/python/benchmark/BUILD

@@ -0,0 +1,24 @@
+# Copyright 2023 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.
+
+# Placeholder for internal Python strict library and test compatibility macro.
+
+package(default_visibility = ["//visibility:public"])
+
+licenses(["notice"])
+
+py_library(
+    name = "benchmark_utils",
+    srcs = ["benchmark_utils.py"]
+)

mediapipe/tasks/python/benchmark/benchmark_utils.py

@@ -0,0 +1,58 @@
+# Copyright 2023 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.
+"""Benchmark utils for MediaPipe Tasks."""
+
+import os
+
+
+def get_test_data_path(test_srcdir, file_or_dirname_path: str) -> str:
+  """Determine the test data path.
+
+  Args:
+      test_srcdir: The path to the test source directory.
+      file_or_dirname_path: The path to the file or directory.
+
+  Returns:
+      The full test data path.
+  """
+  """Returns full test data path."""
+  for directory, subdirs, files in os.walk(test_srcdir):
+    for f in subdirs + files:
+      path = os.path.join(directory, f)
+      if path.endswith(file_or_dirname_path):
+        return path
+  raise ValueError(
+      "No %s in test directory: %s." % (file_or_dirname_path, test_srcdir)
+  )
+
+
+def get_model_path(custom_model, default_model_path):
+  """Determine the model path based on the existence of the custom model.
+
+  Args:
+      custom_model: The path to the custom model provided by the user.
+      default_model_path: The path to the default model.
+
+  Returns:
+      The path to the model to be used.
+  """
+  if custom_model is not None and os.path.exists(custom_model):
+    print(f"Using provided model: {custom_model}")
+    return custom_model
+  else:
+    if custom_model is not None:
+      print(f"Warning: Provided model '{custom_model}' not found. "
+            f"Using default model instead.")
+    print(f"Using default model: {default_model_path}")
+    return default_model_path

mediapipe/tasks/python/benchmark/vision/core/BUILD

@@ -0,0 +1,22 @@
+# Copyright 2023 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.
+
+# Placeholder for internal Python strict library and test compatibility macro.
+
+package(default_visibility = ["//visibility:public"])
+
+py_library(
+    name = "base_vision_benchmark_api",
+    srcs = ["base_vision_benchmark_api.py"]
+)

mediapipe/tasks/python/benchmark/vision/core/__init__.py

@@ -0,0 +1,14 @@
+"""Copyright 2023 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.
+"""

mediapipe/tasks/python/benchmark/vision/core/base_vision_benchmark_api.py

@@ -0,0 +1,44 @@
+# Copyright 2023 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.
+"""MediaPipe vision benchmark base api."""
+import os
+import time
+import numpy as np
+
+VISION_TEST_DATA_DIR = 'mediapipe/tasks/testdata/vision'
+
+
+def nth_percentile(func, image, n_iterations, percentile):
+  """Run a nth percentile benchmark for a given task using the function.
+
+  Args:
+      func: The method associated with a given task used for benchmarking.
+      image: The input MediaPipe Image.
+      n_iterations: Number of iterations to run the benchmark.
+      percentile: Percentage for the percentiles to compute. Values must be
+        between 0 and 100 inclusive.
+
+  Returns:
+    The n-th percentile of the inference times in milliseconds.
+  """
+  inference_times = []
+
+  for _ in range(n_iterations):
+    start_time_ns = time.time_ns()
+    # Run the method for the task (e.g., classify)
+    func(image)
+    end_time_ns = time.time_ns()
+    inference_times.append((end_time_ns - start_time_ns) / 1_000_000)
+
+  return np.percentile(inference_times, percentile)

mediapipe/tasks/python/benchmark/vision/face_aligner/BUILD

@@ -0,0 +1,34 @@
+# Copyright 2022 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.
+
+# Placeholder for internal Python strict library and test compatibility macro.
+
+package(default_visibility = ["//visibility:public"])
+
+py_binary(
+    name = "face_aligner_benchmark",
+    main = "face_aligner_benchmark.py",
+    srcs = ["face_aligner_benchmark.py"],
+    data = [
+        "//mediapipe/tasks/testdata/vision:test_images",
+        "//mediapipe/tasks/testdata/vision:test_models",
+    ],
+    deps = [
+        "//mediapipe/python:_framework_bindings",
+        "//mediapipe/tasks/python/core:base_options",
+        "//mediapipe/tasks/python/vision:face_aligner",
+        "//mediapipe/tasks/python/benchmark:benchmark_utils",
+        "//mediapipe/tasks/python/benchmark/vision/core:base_vision_benchmark_api",
+    ],
+)

mediapipe/tasks/python/benchmark/vision/face_aligner/face_aligner_benchmark.py

@@ -0,0 +1,120 @@
+# Copyright 2023 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.
+"""MediaPipe face aligner benchmark."""
+
+import argparse
+
+from mediapipe.python._framework_bindings import image
+from mediapipe.tasks.python.core import base_options
+from mediapipe.tasks.python.vision import face_aligner
+from mediapipe.tasks.python.benchmark import benchmark_utils
+from mediapipe.tasks.python.benchmark.vision.core import base_vision_benchmark_api
+
+_MODEL_FILE = 'face_landmarker_v2.task'
+_IMAGE_FILE = 'portrait.jpg'
+
+
+def run(
+    model: str,
+    n_iterations: int,
+    delegate: base_options.BaseOptions.Delegate,
+    percentile: float,
+):
+  """Run an face aligner benchmark.
+
+  Args:
+      model: Path to the TFLite model.
+      n_iterations: Number of iterations to run the benchmark.
+      delegate: CPU or GPU delegate for inference.
+      percentile: Percentage for the percentiles to compute. Values must be
+        between 0 and 100 inclusive.
+
+  Returns:
+    The n-th percentile of the inference times.
+  """
+  # Initialize the face aligner
+  default_model_path = benchmark_utils.get_test_data_path(
+    base_vision_benchmark_api.VISION_TEST_DATA_DIR, _MODEL_FILE
+  )
+  model_path = benchmark_utils.get_model_path(model, default_model_path)
+  options = face_aligner.FaceAlignerOptions(
+      base_options=base_options.BaseOptions(
+          model_asset_path=model_path, delegate=delegate
+      )
+  )
+
+  with face_aligner.FaceAligner.create_from_options(options) as aligner:
+    mp_image = image.Image.create_from_file(
+        benchmark_utils.get_test_data_path(
+            base_vision_benchmark_api.VISION_TEST_DATA_DIR, _IMAGE_FILE
+        )
+    )
+    # Run the benchmark and return the nth percentile of the inference times
+    nth_percentile = base_vision_benchmark_api.nth_percentile(
+        aligner.align, mp_image, n_iterations, percentile
+    )
+  return nth_percentile
+
+
+def main():
+  parser = argparse.ArgumentParser(
+      formatter_class=argparse.ArgumentDefaultsHelpFormatter
+  )
+  parser.add_argument(
+      '--model',
+      help='Path to face aligner task.',
+      required=False,
+      default=None,
+  )
+  parser.add_argument(
+      '--iterations',
+      help='Number of iterations for benchmarking.',
+      type=int,
+      default=100,
+  )
+  parser.add_argument(
+      '--percentile',
+      help='Percentile for benchmarking statistics.',
+      type=float,
+      default=95.0,
+  )
+  args = parser.parse_args()
+
+  # Run benchmark on CPU
+  cpu_time = run(
+      args.model,
+      args.iterations,
+      base_options.BaseOptions.Delegate.CPU,
+      args.percentile,
+  )
+  print(
+      f'{args.percentile}th Percentile Inference Time on CPU: '
+      f'{cpu_time:.6f} milliseconds'
+  )
+
+  # Run benchmark on GPU
+  gpu_time = run(
+      args.model,
+      args.iterations,
+      base_options.BaseOptions.Delegate.GPU,
+      args.percentile,
+  )
+  print(
+      f'{args.percentile}th Percentile Inference Time on GPU: '
+      f'{gpu_time:.6f} milliseconds'
+  )
+
+
+if __name__ == '__main__':
+  main()

mediapipe/tasks/python/benchmark/vision/face_detector/BUILD

@@ -0,0 +1,34 @@
+# Copyright 2022 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.
+
+# Placeholder for internal Python strict library and test compatibility macro.
+
+package(default_visibility = ["//visibility:public"])
+
+py_binary(
+    name = "face_detector_benchmark",
+    main = "face_detector_benchmark.py",
+    srcs = ["face_detector_benchmark.py"],
+    data = [
+        "//mediapipe/tasks/testdata/vision:test_images",
+        "//mediapipe/tasks/testdata/vision:test_models",
+    ],
+    deps = [
+        "//mediapipe/python:_framework_bindings",
+        "//mediapipe/tasks/python/core:base_options",
+        "//mediapipe/tasks/python/vision:face_detector",
+        "//mediapipe/tasks/python/benchmark:benchmark_utils",
+        "//mediapipe/tasks/python/benchmark/vision/core:base_vision_benchmark_api",
+    ],
+)

mediapipe/tasks/python/benchmark/vision/face_detector/face_detector_benchmark.py

@@ -0,0 +1,120 @@
+# Copyright 2023 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.
+"""MediaPipe face detector benchmark."""
+
+import argparse
+
+from mediapipe.python._framework_bindings import image
+from mediapipe.tasks.python.core import base_options
+from mediapipe.tasks.python.vision import face_detector
+from mediapipe.tasks.python.benchmark import benchmark_utils
+from mediapipe.tasks.python.benchmark.vision.core import base_vision_benchmark_api
+
+_MODEL_FILE = 'face_detection_short_range.tflite'
+_IMAGE_FILE = 'portrait.jpg'
+
+
+def run(
+    model: str,
+    n_iterations: int,
+    delegate: base_options.BaseOptions.Delegate,
+    percentile: float,
+):
+  """Run an face detector benchmark.
+
+  Args:
+      model: Path to the TFLite model.
+      n_iterations: Number of iterations to run the benchmark.
+      delegate: CPU or GPU delegate for inference.
+      percentile: Percentage for the percentiles to compute. Values must be
+        between 0 and 100 inclusive.
+
+  Returns:
+    The n-th percentile of the inference times.
+  """
+  # Initialize the face detector
+  default_model_path = benchmark_utils.get_test_data_path(
+    base_vision_benchmark_api.VISION_TEST_DATA_DIR, _MODEL_FILE
+  )
+  model_path = benchmark_utils.get_model_path(model, default_model_path)
+  options = face_detector.FaceDetectorOptions(
+      base_options=base_options.BaseOptions(
+          model_asset_path=model_path, delegate=delegate
+      )
+  )
+
+  with face_detector.FaceDetector.create_from_options(options) as detector:
+    mp_image = image.Image.create_from_file(
+        benchmark_utils.get_test_data_path(
+            base_vision_benchmark_api.VISION_TEST_DATA_DIR, _IMAGE_FILE
+        )
+    )
+    # Run the benchmark and return the nth percentile of the inference times
+    nth_percentile = base_vision_benchmark_api.nth_percentile(
+        detector.detect, mp_image, n_iterations, percentile
+    )
+  return nth_percentile
+
+
+def main():
+  parser = argparse.ArgumentParser(
+      formatter_class=argparse.ArgumentDefaultsHelpFormatter
+  )
+  parser.add_argument(
+      '--model',
+      help='Path to face detector task.',
+      required=False,
+      default=None,
+  )
+  parser.add_argument(
+      '--iterations',
+      help='Number of iterations for benchmarking.',
+      type=int,
+      default=100,
+  )
+  parser.add_argument(
+      '--percentile',
+      help='Percentile for benchmarking statistics.',
+      type=float,
+      default=95.0,
+  )
+  args = parser.parse_args()
+
+  # Run benchmark on CPU
+  cpu_time = run(
+      args.model,
+      args.iterations,
+      base_options.BaseOptions.Delegate.CPU,
+      args.percentile,
+  )
+  print(
+      f'{args.percentile}th Percentile Inference Time on CPU: '
+      f'{cpu_time:.6f} milliseconds'
+  )
+
+  # Run benchmark on GPU
+  gpu_time = run(
+      args.model,
+      args.iterations,
+      base_options.BaseOptions.Delegate.GPU,
+      args.percentile,
+  )
+  print(
+      f'{args.percentile}th Percentile Inference Time on GPU: '
+      f'{gpu_time:.6f} milliseconds'
+  )
+
+
+if __name__ == '__main__':
+  main()

mediapipe/tasks/python/benchmark/vision/face_landmarker/BUILD

@@ -0,0 +1,34 @@
+# Copyright 2022 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.
+
+# Placeholder for internal Python strict library and test compatibility macro.
+
+package(default_visibility = ["//visibility:public"])
+
+py_binary(
+    name = "face_landmarker_benchmark",
+    main = "face_landmarker_benchmark.py",
+    srcs = ["face_landmarker_benchmark.py"],
+    data = [
+        "//mediapipe/tasks/testdata/vision:test_images",
+        "//mediapipe/tasks/testdata/vision:test_models",
+    ],
+    deps = [
+        "//mediapipe/python:_framework_bindings",
+        "//mediapipe/tasks/python/core:base_options",
+        "//mediapipe/tasks/python/vision:face_landmarker",
+        "//mediapipe/tasks/python/benchmark:benchmark_utils",
+        "//mediapipe/tasks/python/benchmark/vision/core:base_vision_benchmark_api",
+    ],
+)

mediapipe/tasks/python/benchmark/vision/face_landmarker/face_landmarker_benchmark.py

@@ -0,0 +1,120 @@
+# Copyright 2023 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.
+"""MediaPipe face landmarker benchmark."""
+
+import argparse
+
+from mediapipe.python._framework_bindings import image
+from mediapipe.tasks.python.core import base_options
+from mediapipe.tasks.python.vision import face_landmarker
+from mediapipe.tasks.python.benchmark import benchmark_utils
+from mediapipe.tasks.python.benchmark.vision.core import base_vision_benchmark_api
+
+_MODEL_FILE = 'face_landmarker_v2.task'
+_IMAGE_FILE = 'portrait.jpg'
+
+
+def run(
+    model: str,
+    n_iterations: int,
+    delegate: base_options.BaseOptions.Delegate,
+    percentile: float,
+):
+  """Run an face landmarker benchmark.
+
+  Args:
+      model: Path to the TFLite model.
+      n_iterations: Number of iterations to run the benchmark.
+      delegate: CPU or GPU delegate for inference.
+      percentile: Percentage for the percentiles to compute. Values must be
+        between 0 and 100 inclusive.
+
+  Returns:
+    The n-th percentile of the inference times.
+  """
+  # Initialize the face landmarker
+  default_model_path = benchmark_utils.get_test_data_path(
+    base_vision_benchmark_api.VISION_TEST_DATA_DIR, _MODEL_FILE
+  )
+  model_path = benchmark_utils.get_model_path(model, default_model_path)
+  options = face_landmarker.FaceLandmarkerOptions(
+      base_options=base_options.BaseOptions(
+          model_asset_path=model_path, delegate=delegate
+      )
+  )
+
+  with face_landmarker.FaceLandmarker.create_from_options(options) as landmarker:
+    mp_image = image.Image.create_from_file(
+        benchmark_utils.get_test_data_path(
+            base_vision_benchmark_api.VISION_TEST_DATA_DIR, _IMAGE_FILE
+        )
+    )
+    # Run the benchmark and return the nth percentile of the inference times
+    nth_percentile = base_vision_benchmark_api.nth_percentile(
+        landmarker.detect, mp_image, n_iterations, percentile
+    )
+  return nth_percentile
+
+
+def main():
+  parser = argparse.ArgumentParser(
+      formatter_class=argparse.ArgumentDefaultsHelpFormatter
+  )
+  parser.add_argument(
+      '--model',
+      help='Path to face landmarker task.',
+      required=False,
+      default=None,
+  )
+  parser.add_argument(
+      '--iterations',
+      help='Number of iterations for benchmarking.',
+      type=int,
+      default=100,
+  )
+  parser.add_argument(
+      '--percentile',
+      help='Percentile for benchmarking statistics.',
+      type=float,
+      default=95.0,
+  )
+  args = parser.parse_args()
+
+  # Run benchmark on CPU
+  cpu_time = run(
+      args.model,
+      args.iterations,
+      base_options.BaseOptions.Delegate.CPU,
+      args.percentile,
+  )
+  print(
+      f'{args.percentile}th Percentile Inference Time on CPU: '
+      f'{cpu_time:.6f} milliseconds'
+  )
+
+  # Run benchmark on GPU
+  gpu_time = run(
+      args.model,
+      args.iterations,
+      base_options.BaseOptions.Delegate.GPU,
+      args.percentile,
+  )
+  print(
+      f'{args.percentile}th Percentile Inference Time on GPU: '
+      f'{gpu_time:.6f} milliseconds'
+  )
+
+
+if __name__ == '__main__':
+  main()

mediapipe/tasks/python/benchmark/vision/hand_landmarker/BUILD

@@ -0,0 +1,34 @@
+# Copyright 2022 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.
+
+# Placeholder for internal Python strict library and test compatibility macro.
+
+package(default_visibility = ["//visibility:public"])
+
+py_binary(
+    name = "hand_landmarker_benchmark",
+    main = "hand_landmarker_benchmark.py",
+    srcs = ["hand_landmarker_benchmark.py"],
+    data = [
+        "//mediapipe/tasks/testdata/vision:test_images",
+        "//mediapipe/tasks/testdata/vision:test_models",
+    ],
+    deps = [
+        "//mediapipe/python:_framework_bindings",
+        "//mediapipe/tasks/python/core:base_options",
+        "//mediapipe/tasks/python/vision:hand_landmarker",
+        "//mediapipe/tasks/python/benchmark:benchmark_utils",
+        "//mediapipe/tasks/python/benchmark/vision/core:base_vision_benchmark_api",
+    ],
+)

mediapipe/tasks/python/benchmark/vision/hand_landmarker/hand_landmarker_benchmark.py

@@ -0,0 +1,120 @@
+# Copyright 2023 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.
+"""MediaPipe hand landmarker benchmark."""
+
+import argparse
+
+from mediapipe.python._framework_bindings import image
+from mediapipe.tasks.python.core import base_options
+from mediapipe.tasks.python.vision import hand_landmarker
+from mediapipe.tasks.python.benchmark import benchmark_utils
+from mediapipe.tasks.python.benchmark.vision.core import base_vision_benchmark_api
+
+_MODEL_FILE = 'hand_landmarker.task'
+_IMAGE_FILE = 'thumb_up.jpg'
+
+
+def run(
+    model: str,
+    n_iterations: int,
+    delegate: base_options.BaseOptions.Delegate,
+    percentile: float,
+):
+  """Run an hand landmarker benchmark.
+
+  Args:
+      model: Path to the TFLite model.
+      n_iterations: Number of iterations to run the benchmark.
+      delegate: CPU or GPU delegate for inference.
+      percentile: Percentage for the percentiles to compute. Values must be
+        between 0 and 100 inclusive.
+
+  Returns:
+    The n-th percentile of the inference times.
+  """
+  # Initialize the hand landmarker
+  default_model_path = benchmark_utils.get_test_data_path(
+    base_vision_benchmark_api.VISION_TEST_DATA_DIR, _MODEL_FILE
+  )
+  model_path = benchmark_utils.get_model_path(model, default_model_path)
+  options = hand_landmarker.HandLandmarkerOptions(
+      base_options=base_options.BaseOptions(
+          model_asset_path=model_path, delegate=delegate
+      )
+  )
+
+  with hand_landmarker.HandLandmarker.create_from_options(options) as landmarker:
+    mp_image = image.Image.create_from_file(
+        benchmark_utils.get_test_data_path(
+            base_vision_benchmark_api.VISION_TEST_DATA_DIR, _IMAGE_FILE
+        )
+    )
+    # Run the benchmark and return the nth percentile of the inference times
+    nth_percentile = base_vision_benchmark_api.nth_percentile(
+        landmarker.detect, mp_image, n_iterations, percentile
+    )
+  return nth_percentile
+
+
+def main():
+  parser = argparse.ArgumentParser(
+      formatter_class=argparse.ArgumentDefaultsHelpFormatter
+  )
+  parser.add_argument(
+      '--model',
+      help='Path to hand landmarker task.',
+      required=False,
+      default=None,
+  )
+  parser.add_argument(
+      '--iterations',
+      help='Number of iterations for benchmarking.',
+      type=int,
+      default=100,
+  )
+  parser.add_argument(
+      '--percentile',
+      help='Percentile for benchmarking statistics.',
+      type=float,
+      default=95.0,
+  )
+  args = parser.parse_args()
+
+  # Run benchmark on CPU
+  cpu_time = run(
+      args.model,
+      args.iterations,
+      base_options.BaseOptions.Delegate.CPU,
+      args.percentile,
+  )
+  print(
+      f'{args.percentile}th Percentile Inference Time on CPU: '
+      f'{cpu_time:.6f} milliseconds'
+  )
+
+  # Run benchmark on GPU
+  gpu_time = run(
+      args.model,
+      args.iterations,
+      base_options.BaseOptions.Delegate.GPU,
+      args.percentile,
+  )
+  print(
+      f'{args.percentile}th Percentile Inference Time on GPU: '
+      f'{gpu_time:.6f} milliseconds'
+  )
+
+
+if __name__ == '__main__':
+  main()

mediapipe/tasks/python/benchmark/vision/image_classifier/BUILD

@@ -16,12 +16,19 @@
 
 package(default_visibility = ["//visibility:public"])
 
-py_library(
+py_binary(
     name = "image_classifier_benchmark",
+    main = "image_classifier_benchmark.py",
     srcs = ["image_classifier_benchmark.py"],
+    data = [
+        "//mediapipe/tasks/testdata/vision:test_images",
+        "//mediapipe/tasks/testdata/vision:test_models",
+    ],
     deps = [
         "//mediapipe/python:_framework_bindings",
         "//mediapipe/tasks/python/core:base_options",
         "//mediapipe/tasks/python/vision:image_classifier",
+        "//mediapipe/tasks/python/benchmark:benchmark_utils",
+        "//mediapipe/tasks/python/benchmark/vision/core:base_vision_benchmark_api",
     ],
 )

mediapipe/tasks/python/benchmark/vision/image_classifier/image_classifier_benchmark.py

@@ -14,12 +14,14 @@
 """MediaPipe image classsifier benchmark."""
 
 import argparse
-import time
-import numpy as np
+
 from mediapipe.python._framework_bindings import image
 from mediapipe.tasks.python.core import base_options
 from mediapipe.tasks.python.vision import image_classifier
+from mediapipe.tasks.python.benchmark import benchmark_utils
+from mediapipe.tasks.python.benchmark.vision.core import base_vision_benchmark_api
 
+_MODEL_FILE = 'mobilenet_v2_1.0_224.tflite'
 _IMAGE_FILE = 'burger.jpg'
 
 
@@ -41,27 +43,29 @@ def run(
   Returns:
     The n-th percentile of the inference times.
   """
-  inference_times = []
-
   # Initialize the image classifier
+  default_model_path = benchmark_utils.get_test_data_path(
+      base_vision_benchmark_api.VISION_TEST_DATA_DIR, _MODEL_FILE
+  )
+  model_path = benchmark_utils.get_model_path(model, default_model_path)
   options = image_classifier.ImageClassifierOptions(
       base_options=base_options.BaseOptions(
-          model_asset_path=model, delegate=delegate
+          model_asset_path=model_path, delegate=delegate
       ),
       max_results=1,
   )
-  classifier = image_classifier.ImageClassifier.create_from_options(options)
-  mp_image = image.Image.create_from_file(_IMAGE_FILE)
 
-  for _ in range(n_iterations):
-    start_time_ns = time.time_ns()
-    classifier.classify(mp_image)
-    end_time_ns = time.time_ns()
-    # Convert to milliseconds
-    inference_times.append((end_time_ns - start_time_ns) / 1_000_000)
-
-  classifier.close()
-  return np.percentile(inference_times, percentile)
+  with image_classifier.ImageClassifier.create_from_options(options) as classifier:
+    mp_image = image.Image.create_from_file(
+        benchmark_utils.get_test_data_path(
+            base_vision_benchmark_api.VISION_TEST_DATA_DIR, _IMAGE_FILE
+        )
+    )
+    # Run the benchmark and return the nth percentile of the inference times
+    nth_percentile = base_vision_benchmark_api.nth_percentile(
+        classifier.classify, mp_image, n_iterations, percentile
+    )
+  return nth_percentile
 
 
 def main():
@@ -72,7 +76,7 @@ def main():
       '--model',
       help='Path to image classification model.',
       required=False,
-      default='classifier.tflite',
+      default=None,
   )
   parser.add_argument(
       '--iterations',

mediapipe/tasks/python/benchmark/vision/image_embedder/BUILD

@@ -0,0 +1,34 @@
+# Copyright 2022 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.
+
+# Placeholder for internal Python strict library and test compatibility macro.
+
+package(default_visibility = ["//visibility:public"])
+
+py_binary(
+    name = "image_embedder_benchmark",
+    main = "image_embedder_benchmark.py",
+    srcs = ["image_embedder_benchmark.py"],
+    data = [
+        "//mediapipe/tasks/testdata/vision:test_images",
+        "//mediapipe/tasks/testdata/vision:test_models",
+    ],
+    deps = [
+        "//mediapipe/python:_framework_bindings",
+        "//mediapipe/tasks/python/core:base_options",
+        "//mediapipe/tasks/python/vision:image_embedder",
+        "//mediapipe/tasks/python/benchmark:benchmark_utils",
+        "//mediapipe/tasks/python/benchmark/vision/core:base_vision_benchmark_api",
+    ],
+)

mediapipe/tasks/python/benchmark/vision/image_embedder/image_embedder_benchmark.py

@@ -0,0 +1,120 @@
+# Copyright 2023 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.
+"""MediaPipe image embedder benchmark."""
+
+import argparse
+
+from mediapipe.python._framework_bindings import image
+from mediapipe.tasks.python.core import base_options
+from mediapipe.tasks.python.vision import image_embedder
+from mediapipe.tasks.python.benchmark import benchmark_utils
+from mediapipe.tasks.python.benchmark.vision.core import base_vision_benchmark_api
+
+_MODEL_FILE = 'mobilenet_v3_small_100_224_embedder.tflite'
+_IMAGE_FILE = 'burger.jpg'
+
+
+def run(
+    model: str,
+    n_iterations: int,
+    delegate: base_options.BaseOptions.Delegate,
+    percentile: float,
+):
+  """Run an image embedding extraction benchmark.
+
+  Args:
+      model: Path to the TFLite model.
+      n_iterations: Number of iterations to run the benchmark.
+      delegate: CPU or GPU delegate for inference.
+      percentile: Percentage for the percentiles to compute. Values must be
+        between 0 and 100 inclusive.
+
+  Returns:
+    The n-th percentile of the inference times.
+  """
+  # Initialize the image embedder
+  default_model_path = benchmark_utils.get_test_data_path(
+      base_vision_benchmark_api.VISION_TEST_DATA_DIR, _MODEL_FILE
+  )
+  model_path = benchmark_utils.get_model_path(model, default_model_path)
+  options = image_embedder.ImageEmbedderOptions(
+      base_options=base_options.BaseOptions(
+          model_asset_path=model_path, delegate=delegate
+      )
+  )
+
+  with image_embedder.ImageEmbedder.create_from_options(options) as embedder:
+    mp_image = image.Image.create_from_file(
+        benchmark_utils.get_test_data_path(
+            base_vision_benchmark_api.VISION_TEST_DATA_DIR, _IMAGE_FILE
+        )
+    )
+    # Run the benchmark and return the nth percentile of the inference times
+    nth_percentile = base_vision_benchmark_api.nth_percentile(
+        embedder.embed, mp_image, n_iterations, percentile
+    )
+  return nth_percentile
+
+
+def main():
+  parser = argparse.ArgumentParser(
+      formatter_class=argparse.ArgumentDefaultsHelpFormatter
+  )
+  parser.add_argument(
+      '--model',
+      help='Path to image embedding extraction model.',
+      required=False,
+      default=None,
+  )
+  parser.add_argument(
+      '--iterations',
+      help='Number of iterations for benchmarking.',
+      type=int,
+      default=100,
+  )
+  parser.add_argument(
+      '--percentile',
+      help='Percentile for benchmarking statistics.',
+      type=float,
+      default=95.0,
+  )
+  args = parser.parse_args()
+
+  # Run benchmark on CPU
+  cpu_time = run(
+      args.model,
+      args.iterations,
+      base_options.BaseOptions.Delegate.CPU,
+      args.percentile,
+  )
+  print(
+      f'{args.percentile}th Percentile Inference Time on CPU: '
+      f'{cpu_time:.6f} milliseconds'
+  )
+
+  # Run benchmark on GPU
+  gpu_time = run(
+      args.model,
+      args.iterations,
+      base_options.BaseOptions.Delegate.GPU,
+      args.percentile,
+  )
+  print(
+      f'{args.percentile}th Percentile Inference Time on GPU: '
+      f'{gpu_time:.6f} milliseconds'
+  )
+
+
+if __name__ == '__main__':
+  main()

mediapipe/tasks/python/benchmark/vision/image_segmenter/BUILD

@@ -0,0 +1,34 @@
+# Copyright 2022 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.
+
+# Placeholder for internal Python strict library and test compatibility macro.
+
+package(default_visibility = ["//visibility:public"])
+
+py_binary(
+    name = "image_segmenter_benchmark",
+    main = "image_segmenter_benchmark.py",
+    srcs = ["image_segmenter_benchmark.py"],
+    data = [
+        "//mediapipe/tasks/testdata/vision:test_images",
+        "//mediapipe/tasks/testdata/vision:test_models",
+    ],
+    deps = [
+        "//mediapipe/python:_framework_bindings",
+        "//mediapipe/tasks/python/core:base_options",
+        "//mediapipe/tasks/python/vision:image_segmenter",
+        "//mediapipe/tasks/python/benchmark:benchmark_utils",
+        "//mediapipe/tasks/python/benchmark/vision/core:base_vision_benchmark_api",
+    ],
+)

mediapipe/tasks/python/benchmark/vision/image_segmenter/image_segmenter_benchmark.py

@@ -0,0 +1,121 @@
+# Copyright 2023 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.
+"""MediaPipe image segmenter benchmark."""
+
+import argparse
+
+from mediapipe.python._framework_bindings import image
+from mediapipe.tasks.python.core import base_options
+from mediapipe.tasks.python.vision import image_segmenter
+from mediapipe.tasks.python.benchmark import benchmark_utils
+from mediapipe.tasks.python.benchmark.vision.core import base_vision_benchmark_api
+
+_MODEL_FILE = 'deeplabv3.tflite'
+_IMAGE_FILE = 'segmentation_input_rotation0.jpg'
+
+
+def run(
+    model: str,
+    n_iterations: int,
+    delegate: base_options.BaseOptions.Delegate,
+    percentile: float,
+):
+  """Run an image segmentation benchmark.
+
+  Args:
+      model: Path to the TFLite model.
+      n_iterations: Number of iterations to run the benchmark.
+      delegate: CPU or GPU delegate for inference.
+      percentile: Percentage for the percentiles to compute. Values must be
+        between 0 and 100 inclusive.
+
+  Returns:
+    The n-th percentile of the inference times.
+  """
+  # Initialize the image segmenter
+  default_model_path = benchmark_utils.get_test_data_path(
+      base_vision_benchmark_api.VISION_TEST_DATA_DIR, _MODEL_FILE
+  )
+  model_path = benchmark_utils.get_model_path(model, default_model_path)
+  options = image_segmenter.ImageSegmenterOptions(
+      base_options=base_options.BaseOptions(
+          model_asset_path=model_path, delegate=delegate
+      ),
+      output_confidence_masks=True, output_category_mask=True
+  )
+
+  with image_segmenter.ImageSegmenter.create_from_options(options) as segmenter:
+    mp_image = image.Image.create_from_file(
+        benchmark_utils.get_test_data_path(
+            base_vision_benchmark_api.VISION_TEST_DATA_DIR, _IMAGE_FILE
+        )
+    )
+    # Run the benchmark and return the nth percentile of the inference times
+    nth_percentile = base_vision_benchmark_api.nth_percentile(
+        segmenter.segment, mp_image, n_iterations, percentile
+    )
+  return nth_percentile
+
+
+def main():
+  parser = argparse.ArgumentParser(
+      formatter_class=argparse.ArgumentDefaultsHelpFormatter
+  )
+  parser.add_argument(
+      '--model',
+      help='Path to image segmentation model.',
+      required=False,
+      default=None,
+  )
+  parser.add_argument(
+      '--iterations',
+      help='Number of iterations for benchmarking.',
+      type=int,
+      default=100,
+  )
+  parser.add_argument(
+      '--percentile',
+      help='Percentile for benchmarking statistics.',
+      type=float,
+      default=95.0,
+  )
+  args = parser.parse_args()
+
+  # Run benchmark on CPU
+  cpu_time = run(
+      args.model,
+      args.iterations,
+      base_options.BaseOptions.Delegate.CPU,
+      args.percentile,
+  )
+  print(
+      f'{args.percentile}th Percentile Inference Time on CPU: '
+      f'{cpu_time:.6f} milliseconds'
+  )
+
+  # Run benchmark on GPU
+  gpu_time = run(
+      args.model,
+      args.iterations,
+      base_options.BaseOptions.Delegate.GPU,
+      args.percentile,
+  )
+  print(
+      f'{args.percentile}th Percentile Inference Time on GPU: '
+      f'{gpu_time:.6f} milliseconds'
+  )
+
+
+if __name__ == '__main__':
+  main()

mediapipe/tasks/python/benchmark/vision/interactive_segmenter/BUILD

@@ -0,0 +1,34 @@
+# Copyright 2022 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.
+
+# Placeholder for internal Python strict library and test compatibility macro.
+
+package(default_visibility = ["//visibility:public"])
+
+py_binary(
+    name = "interactive_segmenter_benchmark",
+    main = "interactive_segmenter_benchmark.py",
+    srcs = ["interactive_segmenter_benchmark.py"],
+    data = [
+        "//mediapipe/tasks/testdata/vision:test_images",
+        "//mediapipe/tasks/testdata/vision:test_models",
+    ],
+    deps = [
+        "//mediapipe/python:_framework_bindings",
+        "//mediapipe/tasks/python/core:base_options",
+        "//mediapipe/tasks/python/vision:interactive_segmenter",
+        "//mediapipe/tasks/python/benchmark:benchmark_utils",
+        "//mediapipe/tasks/python/benchmark/vision/core:base_vision_benchmark_api",
+    ],
+)

mediapipe/tasks/python/benchmark/vision/interactive_segmenter/interactive_segmenter_benchmark.py

@@ -0,0 +1,128 @@
+# Copyright 2023 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.
+"""MediaPipe interactive segmenter benchmark."""
+
+from functools import partial
+import argparse
+
+from mediapipe.python._framework_bindings import image
+from mediapipe.tasks.python.core import base_options
+from mediapipe.tasks.python.components.containers import keypoint
+from mediapipe.tasks.python.vision import interactive_segmenter
+from mediapipe.tasks.python.benchmark import benchmark_utils
+from mediapipe.tasks.python.benchmark.vision.core import base_vision_benchmark_api
+
+_MODEL_FILE = 'deeplabv3.tflite'
+_IMAGE_FILE = 'segmentation_input_rotation0.jpg'
+
+
+def run(
+    model: str,
+    n_iterations: int,
+    delegate: base_options.BaseOptions.Delegate,
+    percentile: float,
+):
+  """Run an interactive segmentation benchmark.
+
+  Args:
+      model: Path to the TFLite model.
+      n_iterations: Number of iterations to run the benchmark.
+      delegate: CPU or GPU delegate for inference.
+      percentile: Percentage for the percentiles to compute. Values must be
+        between 0 and 100 inclusive.
+
+  Returns:
+    The n-th percentile of the inference times.
+  """
+  # Initialize the interactive segmenter
+  default_model_path = benchmark_utils.get_test_data_path(
+      base_vision_benchmark_api.VISION_TEST_DATA_DIR, _MODEL_FILE
+  )
+  model_path = benchmark_utils.get_model_path(model, default_model_path)
+
+  options = interactive_segmenter.InteractiveSegmenterOptions(
+      base_options=base_options.BaseOptions(
+          model_asset_path=model_path, delegate=delegate
+      ),
+      output_category_mask=True, output_confidence_masks=False
+  )
+  roi = interactive_segmenter.RegionOfInterest(
+      format=interactive_segmenter.RegionOfInterest.Format.KEYPOINT,
+      keypoint=keypoint.NormalizedKeypoint(0.44, 0.7)
+  )
+
+  with interactive_segmenter.InteractiveSegmenter.create_from_options(options) as segmenter:
+    mp_image = image.Image.create_from_file(
+        benchmark_utils.get_test_data_path(
+            base_vision_benchmark_api.VISION_TEST_DATA_DIR, _IMAGE_FILE
+        )
+    )
+    # Run the benchmark and return the nth percentile of the inference times
+    nth_percentile = base_vision_benchmark_api.nth_percentile(
+        partial(segmenter.segment, roi=roi), mp_image, n_iterations, percentile
+    )
+  return nth_percentile
+
+
+def main():
+  parser = argparse.ArgumentParser(
+      formatter_class=argparse.ArgumentDefaultsHelpFormatter
+  )
+  parser.add_argument(
+      '--model',
+      help='Path to interactive segmentation model.',
+      required=False,
+      default=None,
+  )
+  parser.add_argument(
+      '--iterations',
+      help='Number of iterations for benchmarking.',
+      type=int,
+      default=100,
+  )
+  parser.add_argument(
+      '--percentile',
+      help='Percentile for benchmarking statistics.',
+      type=float,
+      default=95.0,
+  )
+  args = parser.parse_args()
+
+  # Run benchmark on CPU
+  cpu_time = run(
+      args.model,
+      args.iterations,
+      base_options.BaseOptions.Delegate.CPU,
+      args.percentile,
+  )
+  print(
+      f'{args.percentile}th Percentile Inference Time on CPU: '
+      f'{cpu_time:.6f} milliseconds'
+  )
+
+  # Run benchmark on GPU
+  gpu_time = run(
+      args.model,
+      args.iterations,
+      base_options.BaseOptions.Delegate.GPU,
+      args.percentile,
+  )
+  print(
+      f'{args.percentile}th Percentile Inference Time on GPU: '
+      f'{gpu_time:.6f} milliseconds'
+  )
+
+
+if __name__ == '__main__':
+  main()

mediapipe/tasks/python/benchmark/vision/object_detector/BUILD

@@ -0,0 +1,34 @@
+# Copyright 2022 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.
+
+# Placeholder for internal Python strict library and test compatibility macro.
+
+package(default_visibility = ["//visibility:public"])
+
+py_binary(
+    name = "object_detector_benchmark",
+    main = "object_detector_benchmark.py",
+    srcs = ["object_detector_benchmark.py"],
+    data = [
+        "//mediapipe/tasks/testdata/vision:test_images",
+        "//mediapipe/tasks/testdata/vision:test_models",
+    ],
+    deps = [
+        "//mediapipe/python:_framework_bindings",
+        "//mediapipe/tasks/python/core:base_options",
+        "//mediapipe/tasks/python/vision:object_detector",
+        "//mediapipe/tasks/python/benchmark:benchmark_utils",
+        "//mediapipe/tasks/python/benchmark/vision/core:base_vision_benchmark_api",
+    ],
+)

mediapipe/tasks/python/benchmark/vision/object_detector/object_detector_benchmark.py

@@ -0,0 +1,120 @@
+# Copyright 2023 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.
+"""MediaPipe object detector benchmark."""
+
+import argparse
+
+from mediapipe.python._framework_bindings import image
+from mediapipe.tasks.python.core import base_options
+from mediapipe.tasks.python.vision import object_detector
+from mediapipe.tasks.python.benchmark import benchmark_utils
+from mediapipe.tasks.python.benchmark.vision.core import base_vision_benchmark_api
+
+_MODEL_FILE = 'coco_efficientdet_lite0_v1_1.0_quant_2021_09_06.tflite'
+_IMAGE_FILE = 'cats_and_dogs.jpg'
+
+
+def run(
+    model: str,
+    n_iterations: int,
+    delegate: base_options.BaseOptions.Delegate,
+    percentile: float,
+):
+  """Run an object detector benchmark.
+
+  Args:
+      model: Path to the TFLite model.
+      n_iterations: Number of iterations to run the benchmark.
+      delegate: CPU or GPU delegate for inference.
+      percentile: Percentage for the percentiles to compute. Values must be
+        between 0 and 100 inclusive.
+
+  Returns:
+    The n-th percentile of the inference times.
+  """
+  # Initialize the object detector
+  default_model_path = benchmark_utils.get_test_data_path(
+      base_vision_benchmark_api.VISION_TEST_DATA_DIR, _MODEL_FILE
+  )
+  model_path = benchmark_utils.get_model_path(model, default_model_path)
+  options = object_detector.ObjectDetectorOptions(
+      base_options=base_options.BaseOptions(
+          model_asset_path=model_path, delegate=delegate
+      )
+  )
+
+  with object_detector.ObjectDetector.create_from_options(options) as detector:
+    mp_image = image.Image.create_from_file(
+        benchmark_utils.get_test_data_path(
+            base_vision_benchmark_api.VISION_TEST_DATA_DIR, _IMAGE_FILE
+        )
+    )
+    # Run the benchmark and return the nth percentile of the inference times
+    nth_percentile = base_vision_benchmark_api.nth_percentile(
+        detector.detect, mp_image, n_iterations, percentile
+    )
+  return nth_percentile
+
+
+def main():
+  parser = argparse.ArgumentParser(
+      formatter_class=argparse.ArgumentDefaultsHelpFormatter
+  )
+  parser.add_argument(
+      '--model',
+      help='Path to object detector model.',
+      required=False,
+      default=None,
+  )
+  parser.add_argument(
+      '--iterations',
+      help='Number of iterations for benchmarking.',
+      type=int,
+      default=100,
+  )
+  parser.add_argument(
+      '--percentile',
+      help='Percentile for benchmarking statistics.',
+      type=float,
+      default=95.0,
+  )
+  args = parser.parse_args()
+
+  # Run benchmark on CPU
+  cpu_time = run(
+      args.model,
+      args.iterations,
+      base_options.BaseOptions.Delegate.CPU,
+      args.percentile,
+  )
+  print(
+      f'{args.percentile}th Percentile Inference Time on CPU: '
+      f'{cpu_time:.6f} milliseconds'
+  )
+
+  # Run benchmark on GPU
+  gpu_time = run(
+      args.model,
+      args.iterations,
+      base_options.BaseOptions.Delegate.GPU,
+      args.percentile,
+  )
+  print(
+      f'{args.percentile}th Percentile Inference Time on GPU: '
+      f'{gpu_time:.6f} milliseconds'
+  )
+
+
+if __name__ == '__main__':
+  main()

mediapipe/tasks/python/benchmark/vision/pose_landmarker/BUILD

@@ -0,0 +1,34 @@
+# Copyright 2022 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.
+
+# Placeholder for internal Python strict library and test compatibility macro.
+
+package(default_visibility = ["//visibility:public"])
+
+py_binary(
+    name = "pose_landmarker_benchmark",
+    main = "pose_landmarker_benchmark.py",
+    srcs = ["pose_landmarker_benchmark.py"],
+    data = [
+        "//mediapipe/tasks/testdata/vision:test_images",
+        "//mediapipe/tasks/testdata/vision:test_models",
+    ],
+    deps = [
+        "//mediapipe/python:_framework_bindings",
+        "//mediapipe/tasks/python/core:base_options",
+        "//mediapipe/tasks/python/vision:pose_landmarker",
+        "//mediapipe/tasks/python/benchmark:benchmark_utils",
+        "//mediapipe/tasks/python/benchmark/vision/core:base_vision_benchmark_api",
+    ],
+)

mediapipe/tasks/python/benchmark/vision/pose_landmarker/pose_landmarker_benchmark.py

@@ -0,0 +1,120 @@
+# Copyright 2023 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.
+"""MediaPipe pose landmarker benchmark."""
+
+import argparse
+
+from mediapipe.python._framework_bindings import image
+from mediapipe.tasks.python.core import base_options
+from mediapipe.tasks.python.vision import pose_landmarker
+from mediapipe.tasks.python.benchmark import benchmark_utils
+from mediapipe.tasks.python.benchmark.vision.core import base_vision_benchmark_api
+
+_MODEL_FILE = 'pose_landmarker.task'
+_IMAGE_FILE = 'pose.jpg'
+
+
+def run(
+    model: str,
+    n_iterations: int,
+    delegate: base_options.BaseOptions.Delegate,
+    percentile: float,
+):
+  """Run an pose landmarker benchmark.
+
+  Args:
+      model: Path to the TFLite model.
+      n_iterations: Number of iterations to run the benchmark.
+      delegate: CPU or GPU delegate for inference.
+      percentile: Percentage for the percentiles to compute. Values must be
+        between 0 and 100 inclusive.
+
+  Returns:
+    The n-th percentile of the inference times.
+  """
+  # Initialize the pose landmarker
+  default_model_path = benchmark_utils.get_test_data_path(
+    base_vision_benchmark_api.VISION_TEST_DATA_DIR, _MODEL_FILE
+  )
+  model_path = benchmark_utils.get_model_path(model, default_model_path)
+  options = pose_landmarker.PoseLandmarkerOptions(
+      base_options=base_options.BaseOptions(
+          model_asset_path=model_path, delegate=delegate
+      )
+  )
+
+  with pose_landmarker.PoseLandmarker.create_from_options(options) as landmarker:
+    mp_image = image.Image.create_from_file(
+        benchmark_utils.get_test_data_path(
+            base_vision_benchmark_api.VISION_TEST_DATA_DIR, _IMAGE_FILE
+        )
+    )
+    # Run the benchmark and return the nth percentile of the inference times
+    nth_percentile = base_vision_benchmark_api.nth_percentile(
+        landmarker.detect, mp_image, n_iterations, percentile
+    )
+  return nth_percentile
+
+
+def main():
+  parser = argparse.ArgumentParser(
+      formatter_class=argparse.ArgumentDefaultsHelpFormatter
+  )
+  parser.add_argument(
+      '--model',
+      help='Path to pose landmarker task.',
+      required=False,
+      default=None,
+  )
+  parser.add_argument(
+      '--iterations',
+      help='Number of iterations for benchmarking.',
+      type=int,
+      default=100,
+  )
+  parser.add_argument(
+      '--percentile',
+      help='Percentile for benchmarking statistics.',
+      type=float,
+      default=95.0,
+  )
+  args = parser.parse_args()
+
+  # Run benchmark on CPU
+  cpu_time = run(
+      args.model,
+      args.iterations,
+      base_options.BaseOptions.Delegate.CPU,
+      args.percentile,
+  )
+  print(
+      f'{args.percentile}th Percentile Inference Time on CPU: '
+      f'{cpu_time:.6f} milliseconds'
+  )
+
+  # Run benchmark on GPU
+  gpu_time = run(
+      args.model,
+      args.iterations,
+      base_options.BaseOptions.Delegate.GPU,
+      args.percentile,
+  )
+  print(
+      f'{args.percentile}th Percentile Inference Time on GPU: '
+      f'{gpu_time:.6f} milliseconds'
+  )
+
+
+if __name__ == '__main__':
+  main()