From 3fbb2b002bb6e4bc80fa62d7dc4d34d1df1fc287 Mon Sep 17 00:00:00 2001 From: kinaryml Date: Wed, 21 Sep 2022 03:23:04 -0700 Subject: [PATCH] Added image segmenter implementation files --- mediapipe/python/BUILD | 2 + mediapipe/tasks/python/components/BUILD | 28 +++ .../python/components/segmenter_options.py | 78 +++++++ mediapipe/tasks/python/test/vision/BUILD | 38 +++- .../test/vision/image_segmenter_test.py | 118 ++++++++++ mediapipe/tasks/python/vision/BUILD | 19 ++ .../tasks/python/vision/image_segmenter.py | 205 ++++++++++++++++++ 7 files changed, 487 insertions(+), 1 deletion(-) create mode 100644 mediapipe/tasks/python/components/BUILD create mode 100644 mediapipe/tasks/python/components/segmenter_options.py create mode 100644 mediapipe/tasks/python/test/vision/image_segmenter_test.py create mode 100644 mediapipe/tasks/python/vision/image_segmenter.py diff --git a/mediapipe/python/BUILD b/mediapipe/python/BUILD index 3a4a90b44..331ee836c 100644 --- a/mediapipe/python/BUILD +++ b/mediapipe/python/BUILD @@ -86,6 +86,8 @@ cc_library( name = "builtin_task_graphs", deps = [ "//mediapipe/tasks/cc/vision/object_detector:object_detector_graph", + "//mediapipe/tasks/cc/vision/image_classification:image_classifier_graph", + "//mediapipe/tasks/cc/vision/image_segmenter:image_segmenter_graph", ], ) diff --git a/mediapipe/tasks/python/components/BUILD b/mediapipe/tasks/python/components/BUILD new file mode 100644 index 000000000..eb8714a93 --- /dev/null +++ b/mediapipe/tasks/python/components/BUILD @@ -0,0 +1,28 @@ +# Copyright 2022 The MediaPipe Authors. All Rights Reserved. +# +# 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 compatibility macro. + +package(default_visibility = ["//mediapipe/tasks:internal"]) + +licenses(["notice"]) + +py_library( + name = "segmenter_options", + srcs = ["segmenter_options.py"], + deps = [ + "//mediapipe/tasks/cc/components:segmenter_options_py_pb2", + "//mediapipe/tasks/python/core:optional_dependencies", + ], +) diff --git a/mediapipe/tasks/python/components/segmenter_options.py b/mediapipe/tasks/python/components/segmenter_options.py new file mode 100644 index 000000000..5b94a256d --- /dev/null +++ b/mediapipe/tasks/python/components/segmenter_options.py @@ -0,0 +1,78 @@ +# Copyright 2022 The TensorFlow Authors. All Rights Reserved. +# +# 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. +"""Segmenter options data class.""" + +import dataclasses +import enum +from typing import Any, Optional + +from mediapipe.tasks.cc.components import segmenter_options_pb2 +from mediapipe.tasks.python.core.optional_dependencies import doc_controls + +_SegmenterOptionsProto = segmenter_options_pb2.SegmenterOptions + + +class OutputType(enum.Enum): + UNSPECIFIED = 0 + CATEGORY_MASK = 1 + CONFIDENCE_MASK = 2 + + +class Activation(enum.Enum): + NONE = 0 + SIGMOID = 1 + SOFTMAX = 2 + + +@dataclasses.dataclass +class SegmenterOptions: + """Options for segmentation processor. + Attributes: + output_type: The output mask type allows specifying the type of + post-processing to perform on the raw model results. + activation: Activation function to apply to input tensor. + """ + + output_type: Optional[OutputType] = OutputType.CATEGORY_MASK + activation: Optional[Activation] = Activation.NONE + + @doc_controls.do_not_generate_docs + def to_pb2(self) -> _SegmenterOptionsProto: + """Generates a protobuf object to pass to the C++ layer.""" + return _SegmenterOptionsProto( + output_type=self.output_type.value, + activation=self.activation.value + ) + + @classmethod + @doc_controls.do_not_generate_docs + def create_from_pb2( + cls, pb2_obj: _SegmenterOptionsProto) -> "SegmenterOptions": + """Creates a `SegmenterOptions` object from the given protobuf object.""" + return SegmenterOptions( + output_type=OutputType(pb2_obj.output_type), + activation=Activation(pb2_obj.output_type) + ) + + def __eq__(self, other: Any) -> bool: + """Checks if this object is equal to the given object. + Args: + other: The object to be compared with. + Returns: + True if the objects are equal. + """ + if not isinstance(other, SegmenterOptions): + return False + + return self.to_pb2().__eq__(other.to_pb2()) diff --git a/mediapipe/tasks/python/test/vision/BUILD b/mediapipe/tasks/python/test/vision/BUILD index bb495338d..403e00a37 100644 --- a/mediapipe/tasks/python/test/vision/BUILD +++ b/mediapipe/tasks/python/test/vision/BUILD @@ -18,4 +18,40 @@ package(default_visibility = ["//mediapipe/tasks:internal"]) licenses(["notice"]) -# TODO: This test fails in OSS +py_test( + name = "object_detector_test", + srcs = ["object_detector_test.py"], + data = [ + "//mediapipe/tasks/testdata/vision:test_images", + "//mediapipe/tasks/testdata/vision:test_models", + ], + deps = [ + # build rule placeholder: numpy dep, + "//mediapipe/tasks/python/components/containers:bounding_box", + "//mediapipe/tasks/python/components/containers:category", + "//mediapipe/tasks/python/components/containers:detections", + "//mediapipe/tasks/python/core:base_options", + "//mediapipe/tasks/python/test:test_util", + "//mediapipe/tasks/python/vision:object_detector", + "//mediapipe/tasks/python/vision/core:vision_task_running_mode", + "@absl_py//absl/testing:parameterized", + ], +) + +py_test( + name = "image_segmenter_test", + srcs = ["image_segmenter_test.py"], + data = [ + "//mediapipe/tasks/testdata/vision:test_images", + "//mediapipe/tasks/testdata/vision:test_models", + ], + deps = [ + # build rule placeholder: numpy dep, + "//mediapipe/tasks/python/core:base_options", + "//mediapipe/tasks/python/test:test_util", + "//mediapipe/tasks/python/components:segmenter_options", + "//mediapipe/tasks/python/vision:image_segmenter", + "//mediapipe/tasks/python/vision/core:vision_task_running_mode", + "@absl_py//absl/testing:parameterized", + ], +) diff --git a/mediapipe/tasks/python/test/vision/image_segmenter_test.py b/mediapipe/tasks/python/test/vision/image_segmenter_test.py new file mode 100644 index 000000000..704b0fc57 --- /dev/null +++ b/mediapipe/tasks/python/test/vision/image_segmenter_test.py @@ -0,0 +1,118 @@ +# Copyright 2022 The MediaPipe Authors. All Rights Reserved. +# +# 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. +"""Tests for image segmenter.""" + +import enum + +from absl.testing import absltest +from absl.testing import parameterized +import numpy as np + +from mediapipe.python._framework_bindings import image as image_module +from mediapipe.tasks.python.components import segmenter_options +from mediapipe.tasks.python.core import base_options as base_options_module +from mediapipe.tasks.python.test import test_util +from mediapipe.tasks.python.vision import image_segmenter +from mediapipe.tasks.python.vision.core import vision_task_running_mode as running_mode_module + +_BaseOptions = base_options_module.BaseOptions +_Image = image_module.Image +_OutputType = segmenter_options.OutputType +_Activation = segmenter_options.Activation +_ImageSegmenter = image_segmenter.ImageSegmenter +_ImageSegmenterOptions = image_segmenter.ImageSegmenterOptions +_RUNNING_MODE = running_mode_module.VisionTaskRunningMode + +_MODEL_FILE = 'deeplabv3.tflite' +_IMAGE_FILE = 'segmentation_input_rotation0.jpg' +_SEGMENTATION_FILE = 'segmentation_golden_rotation0.png' +_MASK_MAGNIFICATION_FACTOR = 10 +_MATCH_PIXELS_THRESHOLD = 0.01 + + +class ModelFileType(enum.Enum): + FILE_CONTENT = 1 + FILE_NAME = 2 + + +class ImageSegmenterTest(parameterized.TestCase): + + def setUp(self): + super().setUp() + self.test_image = test_util.read_test_image( + test_util.get_test_data_path(_IMAGE_FILE)) + self.model_path = test_util.get_test_data_path(_MODEL_FILE) + + def test_create_from_file_succeeds_with_valid_model_path(self): + # Creates with default option and valid model file successfully. + with _ImageSegmenter.create_from_model_path(self.model_path) as segmenter: + self.assertIsInstance(segmenter, _ImageSegmenter) + + def test_create_from_options_succeeds_with_valid_model_path(self): + # Creates with options containing model file successfully. + base_options = _BaseOptions(file_name=self.model_path) + options = _ImageSegmenterOptions(base_options=base_options) + with _ImageSegmenter.create_from_options(options) as segmenter: + self.assertIsInstance(segmenter, _ImageSegmenter) + + def test_create_from_options_fails_with_invalid_model_path(self): + # Invalid empty model path. + with self.assertRaisesRegex( + ValueError, + r"ExternalFile must specify at least one of 'file_content', " + r"'file_name' or 'file_descriptor_meta'."): + base_options = _BaseOptions(file_name='') + options = _ImageSegmenterOptions(base_options=base_options) + _ImageSegmenter.create_from_options(options) + + def test_create_from_options_succeeds_with_valid_model_content(self): + # Creates with options containing model content successfully. + with open(self.model_path, 'rb') as f: + base_options = _BaseOptions(file_content=f.read()) + options = _ImageSegmenterOptions(base_options=base_options) + segmenter = _ImageSegmenter.create_from_options(options) + self.assertIsInstance(segmenter, _ImageSegmenter) + + @parameterized.parameters( + (ModelFileType.FILE_NAME, 4), + (ModelFileType.FILE_CONTENT, 4)) + def succeeds_with_category_mask(self, model_file_type, max_results): + # Creates segmenter. + if model_file_type is ModelFileType.FILE_NAME: + base_options = _BaseOptions(file_name=self.model_path) + elif model_file_type is ModelFileType.FILE_CONTENT: + with open(self.model_path, 'rb') as f: + model_content = f.read() + base_options = _BaseOptions(file_content=model_content) + else: + # Should never happen + raise ValueError('model_file_type is invalid.') + + options = _ImageSegmenterOptions(base_options=base_options, + output_type=_OutputType.CATEGORY_MASK) + segmenter = _ImageSegmenter.create_from_options(options) + + # Performs image segmentation on the input. + image_result = segmenter.segment(self.test_image) + + # Comparing results. + print(image_result) + + # Closes the segmenter explicitly when the segmenter is not used in + # a context. + segmenter.close() + + +if __name__ == '__main__': + absltest.main() diff --git a/mediapipe/tasks/python/vision/BUILD b/mediapipe/tasks/python/vision/BUILD index 7ff818610..ce59763d6 100644 --- a/mediapipe/tasks/python/vision/BUILD +++ b/mediapipe/tasks/python/vision/BUILD @@ -36,3 +36,22 @@ py_library( "//mediapipe/tasks/python/vision/core:vision_task_running_mode", ], ) + +py_library( + name = "image_segmenter", + srcs = [ + "image_segmenter.py", + ], + deps = [ + "//mediapipe/python:_framework_bindings", + "//mediapipe/python:packet_creator", + "//mediapipe/python:packet_getter", + "//mediapipe/tasks/cc/vision/image_segmenter/proto:image_segmenter_options_py_pb2", + "//mediapipe/tasks/python/components:segmenter_options", + "//mediapipe/tasks/python/core:base_options", + "//mediapipe/tasks/python/core:optional_dependencies", + "//mediapipe/tasks/python/core:task_info", + "//mediapipe/tasks/python/vision/core:base_vision_task_api", + "//mediapipe/tasks/python/vision/core:vision_task_running_mode", + ], +) diff --git a/mediapipe/tasks/python/vision/image_segmenter.py b/mediapipe/tasks/python/vision/image_segmenter.py new file mode 100644 index 000000000..ea40d85c1 --- /dev/null +++ b/mediapipe/tasks/python/vision/image_segmenter.py @@ -0,0 +1,205 @@ +# Copyright 2022 The MediaPipe Authors. All Rights Reserved. +# +# 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 task.""" + +import dataclasses +from typing import Callable, List, Mapping, Optional + +from mediapipe.python import packet_creator +from mediapipe.python import packet_getter +from mediapipe.python._framework_bindings import image as image_module +from mediapipe.python._framework_bindings import packet as packet_module +from mediapipe.python._framework_bindings import task_runner as task_runner_module +from mediapipe.tasks.cc.vision.image_segmenter.proto import image_segmenter_options_pb2 +from mediapipe.tasks.python.components import segmenter_options +from mediapipe.tasks.python.core import base_options as base_options_module +from mediapipe.tasks.python.core import task_info as task_info_module +from mediapipe.tasks.python.core.optional_dependencies import doc_controls +from mediapipe.tasks.python.vision.core import base_vision_task_api +from mediapipe.tasks.python.vision.core import vision_task_running_mode as running_mode_module + +_BaseOptions = base_options_module.BaseOptions +_ImageSegmenterOptionsProto = image_segmenter_options_pb2.ImageSegmenterOptions +_SegmenterOptions = segmenter_options.SegmenterOptions +_RunningMode = running_mode_module.VisionTaskRunningMode +_TaskInfo = task_info_module.TaskInfo +_TaskRunner = task_runner_module.TaskRunner + +_SEGMENTATION_OUT_STREAM_NAME = 'segmented_masks' +_SEGMENTATION_TAG = 'SEGMENTATION' +_GROUPED_SEGMENTATION_TAG = 'GROUPED_SEGMENTATION' +_IMAGE_IN_STREAM_NAME = 'image_in' +_IMAGE_OUT_STREAM_NAME = 'image_out' +_IMAGE_TAG = 'IMAGE' +_TASK_GRAPH_NAME = 'mediapipe.tasks.vision.ImageSegmenterGraph' + + +@dataclasses.dataclass +class ImageSegmenterOptions: + """Options for the image segmenter task. + + Attributes: + base_options: Base options for the image segmenter task. + running_mode: The running mode of the task. Default to the image mode. + Image segmenter task has three running modes: + 1) The image mode for detecting objects on single image inputs. + 2) The video mode for detecting objects on the decoded frames of a video. + 3) The live stream mode for detecting objects on a live stream of input + data, such as from camera. + output_type: Optional output mask type. + activation: Activation function to apply to input tensor. + result_callback: The user-defined result callback for processing live stream + data. The result callback should only be specified when the running mode + is set to the live stream mode. + """ + base_options: _BaseOptions + running_mode: _RunningMode = _RunningMode.IMAGE + output_type: Optional[segmenter_options.OutputType] = segmenter_options.OutputType.CATEGORY_MASK + activation: Optional[segmenter_options.Activation] = segmenter_options.Activation.NONE + result_callback: Optional[ + Callable[[List[image_module.Image], image_module.Image, int], + None]] = None + + @doc_controls.do_not_generate_docs + def to_pb2(self) -> _ImageSegmenterOptionsProto: + """Generates an ImageSegmenterOptions protobuf object.""" + base_options_proto = self.base_options.to_pb2() + base_options_proto.use_stream_mode = False if self.running_mode == _RunningMode.IMAGE else True + + segmenter_options = _SegmenterOptions( + output_type=self.output_type, + activation=self.activation + ) + + return _ImageSegmenterOptionsProto( + base_options=base_options_proto, + segmenter_options=segmenter_options.to_pb2() + ) + + +class ImageSegmenter(base_vision_task_api.BaseVisionTaskApi): + """Class that performs image segmentation on images.""" + + @classmethod + def create_from_model_path(cls, model_path: str) -> 'ImageSegmenter': + """Creates an `ImageSegmenter` object from a TensorFlow Lite model and the default `ImageSegmenterOptions`. + + Note that the created `ImageSegmenter` instance is in image mode, for + performing image segmentation on single image inputs. + + Args: + model_path: Path to the model. + + Returns: + `ImageSegmenter` object that's created from the model file and the default + `ImageSegmenterOptions`. + + Raises: + ValueError: If failed to create `ImageSegmenter` object from the provided + file such as invalid file path. + RuntimeError: If other types of error occurred. + """ + base_options = _BaseOptions(file_name=model_path) + options = ImageSegmenterOptions( + base_options=base_options, running_mode=_RunningMode.IMAGE) + return cls.create_from_options(options) + + @classmethod + def create_from_options(cls, + options: ImageSegmenterOptions) -> 'ImageSegmenter': + """Creates the `ImageSegmenter` object from image segmenter options. + + Args: + options: Options for the image segmenter task. + + Returns: + `ImageSegmenter` object that's created from `options`. + + Raises: + ValueError: If failed to create `ImageSegmenter` object from + `ImageSegmenterOptions` such as missing the model. + RuntimeError: If other types of error occurred. + """ + + def packets_callback(output_packets: Mapping[str, packet_module.Packet]): + if output_packets[_IMAGE_OUT_STREAM_NAME].is_empty(): + return + segmentation_result = packet_getter.get_proto_list( + output_packets[_SEGMENTATION_OUT_STREAM_NAME]) + image = packet_getter.get_image(output_packets[_IMAGE_OUT_STREAM_NAME]) + timestamp = output_packets[_IMAGE_OUT_STREAM_NAME].timestamp + options.result_callback(segmentation_result, image, timestamp) + + task_info = _TaskInfo( + task_graph=_TASK_GRAPH_NAME, + input_streams=[':'.join([_IMAGE_TAG, _IMAGE_IN_STREAM_NAME])], + output_streams=[ + ':'.join([_SEGMENTATION_TAG, _SEGMENTATION_OUT_STREAM_NAME]), + ':'.join([_IMAGE_TAG, _IMAGE_OUT_STREAM_NAME]) + ], + task_options=options) + return cls( + task_info.generate_graph_config( + enable_flow_limiting=options.running_mode == + _RunningMode.LIVE_STREAM), options.running_mode, + packets_callback if options.result_callback else None) + + # TODO: Create an Image class for MediaPipe Tasks. + def segment(self, + image: image_module.Image) -> List[image_module.Image]: + """Performs the actual segmentation task on the provided MediaPipe Image. + + Args: + image: MediaPipe Image. + + Returns: + A segmentation result object that contains a list of segmentation masks + as images. + + Raises: + ValueError: If any of the input arguments is invalid. + RuntimeError: If object detection failed to run. + """ + output_packets = self._process_image_data( + {_IMAGE_IN_STREAM_NAME: packet_creator.create_image(image)}) + segmentation_result = packet_getter.get_proto_list( + output_packets[_SEGMENTATION_OUT_STREAM_NAME]) + return segmentation_result + + # def segment_async(self, image: image_module.Image, timestamp_ms: int) -> None: + # """Sends live image data (an Image with a unique timestamp) to perform image segmentation. + # + # This method will return immediately after the input image is accepted. The + # results will be available via the `result_callback` provided in the + # `ImageSegmenterOptions`. The `segment_async` method is designed to process + # live stream data such as camera input. To lower the overall latency, image + # segmenter may drop the input images if needed. In other words, it's not + # guaranteed to have output per input image. The `result_callback` provides: + # - A segmentation result object that contains a list of segmentation masks + # as images. + # - The input image that the image segmenter runs on. + # - The input timestamp in milliseconds. + # + # Args: + # image: MediaPipe Image. + # timestamp_ms: The timestamp of the input image in milliseconds. + # + # Raises: + # ValueError: If the current input timestamp is smaller than what the object + # detector has already processed. + # """ + # self._send_live_stream_data({ + # _IMAGE_IN_STREAM_NAME: + # packet_creator.create_image(image).at(timestamp_ms) + # })