def __call__(self,
image: Union[Image, np.ndarray, str],
roi: Optional[Rect] = None,
is_right_eye: bool = False) -> IrisResults:
height, width = self.input_shape[1:3]
image_data = image_to_tensor(
image,
roi,
output_size=(width, height),
keep_aspect_ratio=True, # equivalent to scale_mode=FIT
output_range=(0, 1), # see iris_landmark_cpu.pbtxt
flip_horizontal=is_right_eye)
input_data = image_data.tensor_data[np.newaxis]
self.interpreter.set_tensor(self.input_index, input_data)
self.interpreter.invoke()
raw_eye_landmarks = self.interpreter.get_tensor(self.eye_index)
raw_iris_landmarks = self.interpreter.get_tensor(self.iris_index)
height, width = self.input_shape[1:3]
eye_contour = project_landmarks(raw_eye_landmarks,
tensor_size=(width, height),
image_size=image_data.original_size,
padding=image_data.padding,
roi=roi,
flip_horizontal=is_right_eye)
iris_landmarks = project_landmarks(raw_iris_landmarks,
tensor_size=(width, height),
image_size=image_data.original_size,
padding=image_data.padding,
roi=roi,
flip_horizontal=is_right_eye)
return IrisResults(eye_contour, iris_landmarks)
def __call__(
self,
image: Union[Image, np.ndarray, str],
roi: Optional[Rect] = None
) -> List[Landmark]:
"""Run inference and return detections from a given image
Args:
image (Image|ndarray|str): Numpy array of shape
`(height, width, 3)` or PIL Image instance or path to image.
roi (Rect|None): Region within the image that contains a face.
Returns:
(list) List of face landmarks in nromalised coordinates relative to
the input image, i.e. values ranging from [0, 1].
"""
height, width = self.input_shape[1:3]
image_data = image_to_tensor(
image,
roi,
output_size=(width, height),
keep_aspect_ratio=False,
output_range=(0., 1.))
input_data = image_data.tensor_data[np.newaxis]
self.interpreter.set_tensor(self.input_index, input_data)
self.interpreter.invoke()
raw_data = self.interpreter.get_tensor(self.data_index)
raw_face = self.interpreter.get_tensor(self.face_index)
# second tensor contains confidence score for a face detection
face_flag = sigmoid(raw_face).flatten()[-1]
# no data if no face was detected
if face_flag <= DETECTION_THRESHOLD:
return []
# extract and normalise landmark data
height, width = self.input_shape[1:3]
return project_landmarks(raw_data,
tensor_size=(width, height),
image_size=image_data.original_size,
padding=image_data.padding,
roi=roi)
def __call__(self,
image: Union[Image, np.ndarray, str],
roi: Optional[Rect] = None) -> List[Detection]:
"""Run inference and return detections from a given image
Args:
image (Image|ndarray|str): Numpy array of shape
`(height, width, 3)`, PIL Image instance or file name.
roi (Rect|None): Optional region within the image that may
contain faces.
Returns:
(list) List of detection results with relative coordinates.
"""
height, width = self.input_shape[1:3]
image_data = image_to_tensor(image,
roi,
output_size=(width, height),
keep_aspect_ratio=True,
output_range=(-1, 1))
input_data = image_data.tensor_data[np.newaxis]
self.interpreter.set_tensor(self.input_index, input_data)
self.interpreter.invoke()
raw_boxes = self.interpreter.get_tensor(self.bbox_index)
raw_scores = self.interpreter.get_tensor(self.score_index)
boxes = self._decode_boxes(raw_boxes)
scores = self._get_sigmoid_scores(raw_scores)
detections = FaceDetection._convert_to_detections(boxes, scores)
pruned_detections = non_maximum_suppression(detections,
MIN_SUPPRESSION_THRESHOLD,
MIN_SCORE,
weighted=True)
detections = detection_letterbox_removal(pruned_detections,
image_data.padding)
return detections