cassdlcm/mediapipe
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Revised implementation and added more tests

Browse Source
This commit is contained in:
Kinar 2023-12-18 02:47:28 -08:00
parent 88463aeb9e
commit 30e6b766d4
3 changed files with 425 additions and 106 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
mediapipe/tasks/python/test/vision/BUILD
View File

@ -204,13 +204,8 @@ py_test(
],
tags = ["not_run:arm"],
deps = [
"//mediapipe/framework/formats:classification_py_pb2",
"//mediapipe/framework/formats:landmark_py_pb2",
"//mediapipe/tasks/cc/vision/holistic_landmarker/proto:holistic_result_py_pb2",
"//mediapipe/python:_framework_bindings",
"//mediapipe/tasks/python/components/containers:category",
"//mediapipe/tasks/python/components/containers:landmark",
"//mediapipe/tasks/python/components/containers:rect",
"//mediapipe/tasks/python/core:base_options",
"//mediapipe/tasks/python/test:test_utils",
"//mediapipe/tasks/python/vision:holistic_landmarker",

366
mediapipe/tasks/python/test/vision/holistic_landmarker_test.py
View File

@ -14,7 +14,6 @@
"""Tests for holistic landmarker."""
import enum
from typing import List
from unittest import mock
from absl.testing import absltest
@ -22,13 +21,8 @@ from absl.testing import parameterized
import numpy as np
from google.protobuf import text_format
from mediapipe.framework.formats import classification_pb2
from mediapipe.framework.formats import landmark_pb2
from mediapipe.tasks.cc.vision.holistic_landmarker.proto import holistic_result_pb2
from mediapipe.python._framework_bindings import image as image_module
from mediapipe.tasks.python.components.containers import category as category_module
from mediapipe.tasks.python.components.containers import landmark as landmark_module
from mediapipe.tasks.python.components.containers import rect as rect_module
from mediapipe.tasks.python.core import base_options as base_options_module
from mediapipe.tasks.python.test import test_utils
from mediapipe.tasks.python.vision import holistic_landmarker
@ -39,10 +33,6 @@ from mediapipe.tasks.python.vision.core import vision_task_running_mode as runni
HolisticLandmarkerResult = holistic_landmarker.HolisticLandmarkerResult
_HolisticResultProto = holistic_result_pb2.HolisticResult
_BaseOptions = base_options_module.BaseOptions
_Category = category_module.Category
_Rect = rect_module.Rect
_Landmark = landmark_module.Landmark
_NormalizedLandmark = landmark_module.NormalizedLandmark
_Image = image_module.Image
_HolisticLandmarker = holistic_landmarker.HolisticLandmarker
_HolisticLandmarkerOptions = holistic_landmarker.HolisticLandmarkerOptions
@ -53,16 +43,20 @@ _HOLISTIC_LANDMARKER_BUNDLE_ASSET_FILE = 'holistic_landmarker.task'
_POSE_IMAGE = 'male_full_height_hands.jpg'
_CAT_IMAGE = 'cat.jpg'
_EXPECTED_HOLISTIC_RESULT = "male_full_height_hands_result_cpu.pbtxt"
_IMAGE_WIDTH = 638
_IMAGE_HEIGHT = 1000
_LANDMARKS_MARGIN = 0.03
_BLENDSHAPES_MARGIN = 0.13
_VIDEO_LANDMARKS_MARGIN = 0.03
_VIDEO_BLENDSHAPES_MARGIN = 0.31
_LIVE_STREAM_LANDMARKS_MARGIN = 0.03
_LIVE_STREAM_BLENDSHAPES_MARGIN = 0.31
def _get_expected_holistic_landmarker_result(
file_path: str,
) -> HolisticLandmarkerResult:
holistic_result_file_path = test_utils.get_test_data_path(
file_path
)
holistic_result_file_path = test_utils.get_test_data_path(file_path)
with open(holistic_result_file_path, 'rb') as f:
holistic_result_proto = _HolisticResultProto()
# Use this if a .pb file is available.
@ -108,6 +102,7 @@ class HolisticLandmarkerTest(parameterized.TestCase):
for i, elem in enumerate(actual_blendshapes):
self.assertEqual(elem.index, expected_blendshapes[i].index)
self.assertEqual(elem.category_name, expected_blendshapes[i].category_name)
self.assertAlmostEqual(
elem.score,
expected_blendshapes[i].score,
@ -118,7 +113,7 @@ class HolisticLandmarkerTest(parameterized.TestCase):
self,
actual_result: HolisticLandmarkerResult,
expected_result: HolisticLandmarkerResult,
output_segmentation_masks: bool,
output_segmentation_mask: bool,
landmarks_margin: float,
blendshapes_margin: float,
):
@ -134,12 +129,43 @@ class HolisticLandmarkerTest(parameterized.TestCase):
actual_result.face_blendshapes, expected_result.face_blendshapes,
blendshapes_margin
)
if output_segmentation_masks:
self.assertIsInstance(actual_result.segmentation_masks, List)
for _, mask in enumerate(actual_result.segmentation_masks):
self.assertIsInstance(mask, _Image)
if output_segmentation_mask:
self.assertIsInstance(actual_result.segmentation_mask, _Image)
self.assertEqual(actual_result.segmentation_mask.width, _IMAGE_WIDTH)
self.assertEqual(actual_result.segmentation_mask.height, _IMAGE_HEIGHT)
else:
self.assertIsNone(actual_result.segmentation_masks)
self.assertIsNone(actual_result.segmentation_mask)
def test_create_from_file_succeeds_with_valid_model_path(self):
# Creates with default option and valid model file successfully.
with _HolisticLandmarker.create_from_model_path(self.model_path) as landmarker:
self.assertIsInstance(landmarker, _HolisticLandmarker)
def test_create_from_options_succeeds_with_valid_model_path(self):
# Creates with options containing model file successfully.
base_options = _BaseOptions(model_asset_path=self.model_path)
options = _HolisticLandmarkerOptions(base_options=base_options)
with _HolisticLandmarker.create_from_options(options) as landmarker:
self.assertIsInstance(landmarker, _HolisticLandmarker)
def test_create_from_options_fails_with_invalid_model_path(self):
# Invalid empty model path.
with self.assertRaisesRegex(
RuntimeError, 'Unable to open file at /path/to/invalid/model.tflite'
):
base_options = _BaseOptions(
model_asset_path='/path/to/invalid/model.tflite'
)
options = _HolisticLandmarkerOptions(base_options=base_options)
_HolisticLandmarker.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(model_asset_buffer=f.read())
options = _HolisticLandmarkerOptions(base_options=base_options)
landmarker = _HolisticLandmarker.create_from_options(options)
self.assertIsInstance(landmarker, _HolisticLandmarker)
@parameterized.parameters(
(
@ -154,13 +180,25 @@ class HolisticLandmarkerTest(parameterized.TestCase):
False,
_get_expected_holistic_landmarker_result(_EXPECTED_HOLISTIC_RESULT)
),
(
ModelFileType.FILE_NAME,
_HOLISTIC_LANDMARKER_BUNDLE_ASSET_FILE,
True,
_get_expected_holistic_landmarker_result(_EXPECTED_HOLISTIC_RESULT)
),
(
ModelFileType.FILE_CONTENT,
_HOLISTIC_LANDMARKER_BUNDLE_ASSET_FILE,
True,
_get_expected_holistic_landmarker_result(_EXPECTED_HOLISTIC_RESULT)
),
)
def test_detect(
self,
model_file_type,
model_name,
output_segmentation_masks,
expected_holistic_landmarker_result: HolisticLandmarkerResult
output_segmentation_mask,
expected_holistic_landmarker_result
):
# Creates holistic landmarker.
model_path = test_utils.get_test_data_path(model_name)
@ -178,7 +216,7 @@ class HolisticLandmarkerTest(parameterized.TestCase):
base_options=base_options,
output_face_blendshapes=True
if expected_holistic_landmarker_result.face_blendshapes else False,
output_segmentation_masks=output_segmentation_masks,
output_segmentation_mask=output_segmentation_mask,
)
landmarker = _HolisticLandmarker.create_from_options(options)
@ -186,12 +224,294 @@ class HolisticLandmarkerTest(parameterized.TestCase):
detection_result = landmarker.detect(self.test_image)
self._expect_holistic_landmarker_results_correct(
detection_result, expected_holistic_landmarker_result,
output_segmentation_masks, _LANDMARKS_MARGIN, _BLENDSHAPES_MARGIN
output_segmentation_mask, _LANDMARKS_MARGIN, _BLENDSHAPES_MARGIN
)
# Closes the holistic landmarker explicitly when the holistic landmarker is
# not used in a context.
landmarker.close()
@parameterized.parameters(
(
ModelFileType.FILE_NAME,
_HOLISTIC_LANDMARKER_BUNDLE_ASSET_FILE,
False,
_get_expected_holistic_landmarker_result(_EXPECTED_HOLISTIC_RESULT)
),
(
ModelFileType.FILE_CONTENT,
_HOLISTIC_LANDMARKER_BUNDLE_ASSET_FILE,
True,
_get_expected_holistic_landmarker_result(_EXPECTED_HOLISTIC_RESULT)
),
)
def test_detect_in_context(
self,
model_file_type,
model_name,
output_segmentation_mask,
expected_holistic_landmarker_result
):
# Creates holistic landmarker.
model_path = test_utils.get_test_data_path(model_name)
if model_file_type is ModelFileType.FILE_NAME:
base_options = _BaseOptions(model_asset_path=model_path)
elif model_file_type is ModelFileType.FILE_CONTENT:
with open(model_path, 'rb') as f:
model_content = f.read()
base_options = _BaseOptions(model_asset_buffer=model_content)
else:
# Should never happen
raise ValueError('model_file_type is invalid.')
options = _HolisticLandmarkerOptions(
base_options=base_options,
output_face_blendshapes=True
if expected_holistic_landmarker_result.face_blendshapes else False,
output_segmentation_mask=output_segmentation_mask,
)
with _HolisticLandmarker.create_from_options(options) as landmarker:
# Performs holistic landmarks detection on the input.
detection_result = landmarker.detect(self.test_image)
self._expect_holistic_landmarker_results_correct(
detection_result, expected_holistic_landmarker_result,
output_segmentation_mask, _LANDMARKS_MARGIN, _BLENDSHAPES_MARGIN
)
def test_empty_detection_outputs(self):
options = _HolisticLandmarkerOptions(
base_options=_BaseOptions(model_asset_path=self.model_path)
)
with _HolisticLandmarker.create_from_options(options) as landmarker:
# Load the cat image.
cat_test_image = _Image.create_from_file(
test_utils.get_test_data_path(_CAT_IMAGE)
)
# Performs holistic landmarks detection on the input.
detection_result = landmarker.detect(cat_test_image)
self.assertEmpty(detection_result.face_landmarks)
self.assertEmpty(detection_result.pose_landmarks)
self.assertEmpty(detection_result.pose_world_landmarks)
self.assertEmpty(detection_result.left_hand_landmarks)
self.assertEmpty(detection_result.left_hand_world_landmarks)
self.assertEmpty(detection_result.right_hand_landmarks)
self.assertEmpty(detection_result.right_hand_world_landmarks)
self.assertIsNone(detection_result.face_blendshapes)
self.assertIsNone(detection_result.segmentation_mask)
def test_missing_result_callback(self):
options = _HolisticLandmarkerOptions(
base_options=_BaseOptions(model_asset_path=self.model_path),
running_mode=_RUNNING_MODE.LIVE_STREAM,
)
with self.assertRaisesRegex(
ValueError, r'result callback must be provided'
):
with _HolisticLandmarker.create_from_options(options) as unused_landmarker:
pass
@parameterized.parameters((_RUNNING_MODE.IMAGE), (_RUNNING_MODE.VIDEO))
def test_illegal_result_callback(self, running_mode):
options = _HolisticLandmarkerOptions(
base_options=_BaseOptions(model_asset_path=self.model_path),
running_mode=running_mode,
result_callback=mock.MagicMock(),
)
with self.assertRaisesRegex(
ValueError, r'result callback should not be provided'
):
with _HolisticLandmarker.create_from_options(options) as unused_landmarker:
pass
def test_calling_detect_for_video_in_image_mode(self):
options = _HolisticLandmarkerOptions(
base_options=_BaseOptions(model_asset_path=self.model_path),
running_mode=_RUNNING_MODE.IMAGE,
)
with _HolisticLandmarker.create_from_options(options) as landmarker:
with self.assertRaisesRegex(
ValueError, r'not initialized with the video mode'
):
landmarker.detect_for_video(self.test_image, 0)
def test_calling_detect_async_in_image_mode(self):
options = _HolisticLandmarkerOptions(
base_options=_BaseOptions(model_asset_path=self.model_path),
running_mode=_RUNNING_MODE.IMAGE,
)
with _HolisticLandmarker.create_from_options(options) as landmarker:
with self.assertRaisesRegex(
ValueError, r'not initialized with the live stream mode'
):
landmarker.detect_async(self.test_image, 0)
def test_calling_detect_in_video_mode(self):
options = _HolisticLandmarkerOptions(
base_options=_BaseOptions(model_asset_path=self.model_path),
running_mode=_RUNNING_MODE.VIDEO,
)
with _HolisticLandmarker.create_from_options(options) as landmarker:
with self.assertRaisesRegex(
ValueError, r'not initialized with the image mode'
):
landmarker.detect(self.test_image)
def test_calling_detect_async_in_video_mode(self):
options = _HolisticLandmarkerOptions(
base_options=_BaseOptions(model_asset_path=self.model_path),
running_mode=_RUNNING_MODE.VIDEO,
)
with _HolisticLandmarker.create_from_options(options) as landmarker:
with self.assertRaisesRegex(
ValueError, r'not initialized with the live stream mode'
):
landmarker.detect_async(self.test_image, 0)
def test_detect_for_video_with_out_of_order_timestamp(self):
options = _HolisticLandmarkerOptions(
base_options=_BaseOptions(model_asset_path=self.model_path),
running_mode=_RUNNING_MODE.VIDEO,
)
with _HolisticLandmarker.create_from_options(options) as landmarker:
unused_result = landmarker.detect_for_video(self.test_image, 1)
with self.assertRaisesRegex(
ValueError, r'Input timestamp must be monotonically increasing'
):
landmarker.detect_for_video(self.test_image, 0)
@parameterized.parameters(
(
_HOLISTIC_LANDMARKER_BUNDLE_ASSET_FILE,
False,
_get_expected_holistic_landmarker_result(_EXPECTED_HOLISTIC_RESULT)
),
(
_HOLISTIC_LANDMARKER_BUNDLE_ASSET_FILE,
True,
_get_expected_holistic_landmarker_result(_EXPECTED_HOLISTIC_RESULT)
),
)
def test_detect_for_video(
self,
model_name,
output_segmentation_mask,
expected_holistic_landmarker_result
):
# Creates holistic landmarker.
model_path = test_utils.get_test_data_path(model_name)
base_options = _BaseOptions(model_asset_path=model_path)
options = _HolisticLandmarkerOptions(
base_options=base_options,
running_mode=_RUNNING_MODE.VIDEO,
output_face_blendshapes=True
if expected_holistic_landmarker_result.face_blendshapes else False,
output_segmentation_mask=output_segmentation_mask,
)
with _HolisticLandmarker.create_from_options(options) as landmarker:
for timestamp in range(0, 300, 30):
# Performs holistic landmarks detection on the input.
detection_result = landmarker.detect_for_video(
self.test_image, timestamp
)
# Comparing results.
self._expect_holistic_landmarker_results_correct(
detection_result, expected_holistic_landmarker_result,
output_segmentation_mask,
_VIDEO_LANDMARKS_MARGIN, _VIDEO_BLENDSHAPES_MARGIN
)
def test_calling_detect_in_live_stream_mode(self):
options = _HolisticLandmarkerOptions(
base_options=_BaseOptions(model_asset_path=self.model_path),
running_mode=_RUNNING_MODE.LIVE_STREAM,
result_callback=mock.MagicMock(),
)
with _HolisticLandmarker.create_from_options(options) as landmarker:
with self.assertRaisesRegex(
ValueError, r'not initialized with the image mode'
):
landmarker.detect(self.test_image)
def test_calling_detect_for_video_in_live_stream_mode(self):
options = _HolisticLandmarkerOptions(
base_options=_BaseOptions(model_asset_path=self.model_path),
running_mode=_RUNNING_MODE.LIVE_STREAM,
result_callback=mock.MagicMock(),
)
with _HolisticLandmarker.create_from_options(options) as landmarker:
with self.assertRaisesRegex(
ValueError, r'not initialized with the video mode'
):
landmarker.detect_for_video(self.test_image, 0)
def test_detect_async_calls_with_illegal_timestamp(self):
options = _HolisticLandmarkerOptions(
base_options=_BaseOptions(model_asset_path=self.model_path),
running_mode=_RUNNING_MODE.LIVE_STREAM,
result_callback=mock.MagicMock(),
)
with _HolisticLandmarker.create_from_options(options) as landmarker:
landmarker.detect_async(self.test_image, 100)
with self.assertRaisesRegex(
ValueError, r'Input timestamp must be monotonically increasing'
):
landmarker.detect_async(self.test_image, 0)
@parameterized.parameters(
(
_POSE_IMAGE,
_HOLISTIC_LANDMARKER_BUNDLE_ASSET_FILE,
False,
_get_expected_holistic_landmarker_result(_EXPECTED_HOLISTIC_RESULT)
),
(
_POSE_IMAGE,
_HOLISTIC_LANDMARKER_BUNDLE_ASSET_FILE,
True,
_get_expected_holistic_landmarker_result(_EXPECTED_HOLISTIC_RESULT)
),
)
def test_detect_async_calls(
self,
image_path,
model_name,
output_segmentation_mask,
expected_holistic_landmarker_result
):
test_image = _Image.create_from_file(
test_utils.get_test_data_path(image_path)
)
observed_timestamp_ms = -1
def check_result(
result: HolisticLandmarkerResult, output_image: _Image, timestamp_ms: int
):
# Comparing results.
self._expect_holistic_landmarker_results_correct(
result, expected_holistic_landmarker_result,
output_segmentation_mask,
_LIVE_STREAM_LANDMARKS_MARGIN, _LIVE_STREAM_BLENDSHAPES_MARGIN
)
self.assertTrue(
np.array_equal(output_image.numpy_view(), test_image.numpy_view())
)
self.assertLess(observed_timestamp_ms, timestamp_ms)
self.observed_timestamp_ms = timestamp_ms
model_path = test_utils.get_test_data_path(model_name)
options = _HolisticLandmarkerOptions(
base_options=_BaseOptions(model_asset_path=model_path),
running_mode=_RUNNING_MODE.LIVE_STREAM,
output_face_blendshapes=True
if expected_holistic_landmarker_result.face_blendshapes else False,
output_segmentation_mask=output_segmentation_mask,
result_callback=check_result,
)
with _HolisticLandmarker.create_from_options(options) as landmarker:
for timestamp in range(0, 300, 30):
landmarker.detect_async(test_image, timestamp)
if __name__ == '__main__':
absltest.main()

24
mediapipe/tasks/python/vision/holistic_landmarker.py
View File

@ -51,7 +51,7 @@ _POSE_LANDMARKS_TAG_NAME = "POSE_LANDMARKS"
_POSE_WORLD_LANDMARKS_STREAM_NAME = "pose_world_landmarks"
_POSE_WORLD_LANDMARKS_TAG = "POSE_WORLD_LANDMARKS"
_POSE_SEGMENTATION_MASK_STREAM_NAME = "pose_segmentation_mask"
_POSE_SEGMENTATION_MASK_TAG = "pose_segmentation_mask"
_POSE_SEGMENTATION_MASK_TAG = "POSE_SEGMENTATION_MASK"
_FACE_LANDMARKS_STREAM_NAME = "face_landmarks"
_FACE_LANDMARKS_TAG = "FACE_LANDMARKS"
_FACE_BLENDSHAPES_STREAM_NAME = "extra_blendshapes"
@ -84,7 +84,7 @@ class HolisticLandmarkerResult:
right_hand_landmarks: List[landmark_module.NormalizedLandmark]
right_hand_world_landmarks: List[landmark_module.Landmark]
face_blendshapes: Optional[List[category_module.Category]] = None
segmentation_masks: Optional[List[image_module.Image]] = None
segmentation_mask: Optional[image_module.Image] = None
@classmethod
@doc_controls.do_not_generate_docs
@ -237,7 +237,7 @@ def _build_landmarker_result(
)
if _POSE_SEGMENTATION_MASK_STREAM_NAME in output_packets:
holistic_landmarker_result.segmentation_masks = packet_getter.get_image_list(
holistic_landmarker_result.segmentation_mask = packet_getter.get_image(
output_packets[_POSE_SEGMENTATION_MASK_STREAM_NAME]
)
@ -273,7 +273,7 @@ class HolisticLandmarkerOptions:
landmark detection to be considered successful.
output_face_blendshapes: Whether HolisticLandmarker outputs face blendshapes
classification. Face blendshapes are used for rendering the 3D face model.
output_segmentation_masks: whether to output segmentation masks.
output_segmentation_mask: whether to output segmentation masks.
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.
@ -290,7 +290,7 @@ class HolisticLandmarkerOptions:
min_pose_landmarks_confidence: float = 0.5
min_hand_landmarks_confidence: float = 0.5
output_face_blendshapes: bool = False
output_segmentation_masks: bool = False
output_segmentation_mask: bool = False
result_callback: Optional[
Callable[[HolisticLandmarkerResult, image_module.Image, int], None]
] = None
@ -411,18 +411,22 @@ class HolisticLandmarker(base_vision_task_api.BaseVisionTaskApi):
),
':'.join([_LEFT_HAND_LANDMARKS_TAG, _LEFT_HAND_LANDMARKS_STREAM_NAME]),
':'.join(
[_LEFT_HAND_WORLD_LANDMARKS_TAG, _LEFT_HAND_WORLD_LANDMARKS_STREAM_NAME]
[_LEFT_HAND_WORLD_LANDMARKS_TAG,
_LEFT_HAND_WORLD_LANDMARKS_STREAM_NAME]
),
':'.join([_RIGHT_HAND_LANDMARKS_TAG, _RIGHT_HAND_LANDMARKS_STREAM_NAME]),
':'.join([_RIGHT_HAND_LANDMARKS_TAG,
_RIGHT_HAND_LANDMARKS_STREAM_NAME]),
':'.join(
[_RIGHT_HAND_WORLD_LANDMARKS_TAG, _RIGHT_HAND_WORLD_LANDMARKS_STREAM_NAME]
[_RIGHT_HAND_WORLD_LANDMARKS_TAG,
_RIGHT_HAND_WORLD_LANDMARKS_STREAM_NAME]
),
':'.join([_IMAGE_TAG, _IMAGE_OUT_STREAM_NAME]),
]
if options.output_segmentation_masks:
if options.output_segmentation_mask:
output_streams.append(
':'.join([_POSE_SEGMENTATION_MASK_TAG, _POSE_SEGMENTATION_MASK_STREAM_NAME])
':'.join([_POSE_SEGMENTATION_MASK_TAG,
_POSE_SEGMENTATION_MASK_STREAM_NAME])
)
if options.output_face_blendshapes: