def preprocess(self, data):
preprocessed_data = []
data_infos = []
for img in data:
img = np.array(img)[:, :, ::-1]
height, width, _ = img.shape
if self.input_height / self.input_width < height / width:
scale = self.input_height / height
else:
scale = self.input_width / width
scaled_img = cv2.resize(
img, (0, 0), fx=scale, fy=scale, interpolation=cv2.INTER_CUBIC,
)
padded_img, pad = pad_img(
scaled_img, (0, 0, 0), [self.input_height, self.input_width],
)
padded_img = padded_img.transpose((2, 0, 1))
padded_img = padded_img[np.newaxis].astype(np.float32)
preprocessed_data.append(padded_img)
data_infos.append(
DataInfo(
scales=(scale, scale),
pads=tuple(pad),
original_width=width,
original_height=height,
),
)
return [preprocessed_data, data_infos]
def preprocess(self, data):
face_bboxes = self.get_first_stage_result(data)
preprocessed_data = []
preprocessed_bboxes = []
if face_bboxes == [[]]:
return None
for i, img in enumerate(data):
areas = [(bbox.y2 - bbox.y1) * (bbox.x2 - bbox.x1)
for bbox in face_bboxes[i]]
max_area = max(areas)
preprocessed_img = []
img_bboxes = []
img = np.array(img)[:, :, ::-1]
for j, face_bbox in enumerate(face_bboxes[i]):
if areas[j] < self.area_threshold * max_area:
continue
if face_bbox.x2 - face_bbox.x1 < self.input_width:
continue
cropped_face = img[face_bbox.y1:face_bbox.y2,
face_bbox.x1:face_bbox.x2, ]
height, width, _ = cropped_face.shape
if self.input_height / self.input_width < height / width:
scale = self.input_height / height
else:
scale = self.input_width / width
scaled_img = cv2.resize(
cropped_face,
(0, 0),
fx=scale,
fy=scale,
interpolation=cv2.INTER_CUBIC,
)
padded_img, pad = pad_img(
scaled_img,
(0, 0, 0),
[self.input_height, self.input_width],
)
padded_img = padded_img.transpose((2, 0, 1))
padded_img = padded_img[np.newaxis].astype(np.float32)
preprocessed_img.append(padded_img)
img_bboxes.append(face_bbox)
preprocessed_data.append(preprocessed_img)
preprocessed_bboxes.append(img_bboxes)
return [preprocessed_data, preprocessed_bboxes]
def preprocess(self, data):
face_bboxes = self.get_first_stage_result(data)
preprocessed_data = []
preprocessed_bboxes = []
if face_bboxes == [[]]:
return None
for i, img in enumerate(data):
areas = [
(bbox.y2 - bbox.y1) * (bbox.x2 - bbox.x1) for bbox in face_bboxes[i]
]
max_area = max(areas)
preprocessed_img = []
img_bboxes = []
img = np.array(img)[:, :, ::-1]
for j, face_bbox in enumerate(face_bboxes[i]):
if areas[j] < self.area_threshold * max_area:
continue
# apply padding to face bbox to increase quality of facial landmark model
face_bbox_width = face_bbox.x2 - face_bbox.x1
face_bbox_hight = face_bbox.y2 - face_bbox.y1
face_bbox.x1 = int(
np.clip(
face_bbox.x1 - face_bbox_width * self.face_bbox_pad_percent,
0,
img.shape[1],
),
)
face_bbox.y2 = int(
np.clip(
face_bbox.y2 + face_bbox_hight * self.face_bbox_pad_percent,
0,
img.shape[0],
),
)
face_bbox.x2 = int(
np.clip(
face_bbox.x2 + face_bbox_width * self.face_bbox_pad_percent,
0,
img.shape[1],
),
)
cropped_face = img[
face_bbox.y1 : face_bbox.y2, face_bbox.x1 : face_bbox.x2,
]
height, width, _ = cropped_face.shape
if self.input_height / self.input_width < height / width:
scale = self.input_height / height
else:
scale = self.input_width / width
scaled_img = cv2.resize(
cropped_face,
(0, 0),
fx=scale,
fy=scale,
interpolation=cv2.INTER_CUBIC,
)
padded_img, pad = pad_img(
scaled_img, (0, 0, 0), [self.input_height, self.input_width],
)
padded_img = padded_img.transpose((2, 0, 1))
padded_img = padded_img[np.newaxis].astype(np.float32)
preprocessed_img.append(padded_img)
img_bboxes.append(face_bbox)
preprocessed_data.append(preprocessed_img)
preprocessed_bboxes.append(img_bboxes)
return [preprocessed_data, preprocessed_bboxes]
