async def get(self, img, extract_embedding: bool = True, extract_ga: bool = True,
return_face_data: bool = True, max_size: List[int] = None, threshold: float = 0.6):
ts = time.time()
t0 = time.time()
# If detector has input_shape attribute, use it instead of provided value
try:
max_size = self.det_model.retina.input_shape[2:][::-1]
except:
pass
img = ImageData(img, max_size=max_size)
img.resize_image(mode='pad')
t1 = time.time()
logging.debug(f'Preparing image took: {t1 - t0}')
t0 = time.time()
boxes, probs, landmarks, mask_probs = self.det_model.detect(img.transformed_image, threshold=threshold)
t1 = time.time()
logging.debug(f'Detection took: {t1 - t0}')
faces = []
await asyncio.sleep(0)
if not isinstance(boxes, type(None)):
t0 = time.time()
for i in range(len(boxes)):
# Translate points to original image size
bbox = self.reproject_points(boxes[i], img.scale_factor)
landmark = self.reproject_points(landmarks[i], img.scale_factor)
det_score = probs[i]
if not isinstance(mask_probs, type(None)):
mask_prob = mask_probs[i]
else:
mask_prob = None
# Crop faces from original image instead of resized to improve quality
_crop = face_align.norm_crop(img.orig_image, landmark=landmark)
face = Face(bbox=bbox, landmark=landmark, det_score=det_score,
num_det=i, scale=img.scale_factor, mask_prob=mask_prob, facedata=_crop)
faces.append(face)
t1 = time.time()
logging.debug(f'Cropping {len(boxes)} faces took: {t1 - t0}')
# Process detected faces
faces = [e for e in self.process_faces(faces, extract_embedding=extract_embedding,
extract_ga=extract_ga, return_face_data=return_face_data)]
tf = time.time()
logging.debug(f'Full processing took: {tf - ts}')
return faces
def detect(self, image):
image = ImageData(image, self.input_shape)
image.resize_image(mode='pad')
dets, lmks = self.detector.detect(image.transformed_image,
threshold=self.fd_conf)
DETS, LMKS = [], []
for i in range(len(dets)):
face = self.faceCrop(image, dets[i])
if min(face[2], face[3]) > 0:
DETS.append(face)
LMKS.append(lmks[i])
return np.array(DETS), np.array(LMKS)
def face_align(src_img, fd, args):
src_image = ImageData(src_img, fd.input_shape)
src_image.resize_image(mode='pad')
dets, lmks = fd.detector.detect(src_image.transformed_image,
threshold=args.fd_conf)
if dets.shape[0] > 0:
bindex = fd.get_max_face(dets)
landmark = lmks[bindex]
bbox = dets[bindex, :4].astype(np.int)
image_size = '%d,%d' % (args.fr_input_size[1], args.fr_input_size[0])
warped = preprocess(src_image.transformed_image,
bbox,
landmark,
image_size=image_size)
return warped
else:
return None
async def get(self,
img,
extract_embedding: bool = True,
extract_ga: bool = True,
return_face_data: bool = True,
max_size: List[int] = None,
threshold: float = 0.6):
ts = time.time()
t0 = time.time()
# If detector has input_shape attribute, use it instead of provided value
try:
max_size = self.det_model.retina.input_shape[2:][::-1]
except:
pass
img = ImageData(img, max_size=max_size)
img.resize_image(mode='pad')
t1 = time.time()
logging.debug(f'Preparing image took: {t1 - t0}')
t0 = time.time()
boxes, probs, landmarks = self.det_model.detect(img.transformed_image,
threshold=threshold)
t1 = time.time()
logging.debug(f'Detection took: {t1 - t0}')
await asyncio.sleep(0)
ret = []
if not isinstance(boxes, type(None)):
t0 = time.time()
for i in range(len(boxes)):
# Translate points to original image size
bbox = self.reproject_points(boxes[i], img.scale_factor)
landmark = self.reproject_points(landmarks[i],
img.scale_factor)
det_score = probs[i]
# Crop faces from original image instead of resized to improve quality
_crop = face_align.norm_crop(img.orig_image, landmark=landmark)
facedata = None
embedding = None
embedding_norm = None
normed_embedding = None
gender = None
age = None
if return_face_data:
facedata = _crop
if extract_embedding and self.rec_model:
embedding = self.rec_model.get_embedding(_crop).flatten()
embedding_norm = norm(embedding)
normed_embedding = embedding / embedding_norm
if self.ga_model:
if extract_ga:
gender, age = self.ga_model.get(_crop)
gender = int(gender)
face = Face(bbox=bbox,
landmark=landmark,
det_score=det_score,
embedding=embedding,
gender=gender,
age=age,
normed_embedding=normed_embedding,
embedding_norm=embedding_norm,
facedata=facedata,
num_det=i,
scale=img.scale_factor)
ret.append(face)
t1 = time.time()
logging.debug(f'Embedding {len(boxes)} faces took: {t1 - t0}')
await asyncio.sleep(0)
tf = time.time()
logging.debug(f'Full processing took: {tf - ts}')
return ret
backend = 'trt'
im_size = [640, 480]
detector = get_model(model_name, backend, im_size=im_size, root_dir='/models')
detector.prepare(nms=0.35)
# Since dynamic shapes are not yet implemented, images during inference
# must be resized to exact shape which was used for TRT building.
# To ensure this we read expected input shape from engine itself.
input_shape = detector.input_shape[2:][::-1]
logging.info(f"Images will be resized to WxH : {input_shape}")
tt0 = time.time()
image = cv2.imread('test_images2/china.jpg', cv2.IMREAD_COLOR)
image = ImageData(image, input_shape)
image.resize_image(mode='pad')
tt1 = time.time()
print(f"Preparing image took: {tt1 - tt0}")
# Since we are using Numba JIT compilation, first run will cause
# compilation on NMS module
logging.info('Warming up NMS jit...')
detections = detector.detect(image.transformed_image, threshold=0.1)
logging.info('Warming up complete!')
det_count = 0
def get(self, img, extract_embedding: bool = True, extract_ga: bool = True,
return_face_data: bool = True, max_size: List[int] = None, threshold: float = 0.6):
ts = time.time()
t0 = time.time()
# If detector has input_shape attribute, use it instead of provided value
try:
max_size = self.det_model.retina.input_shape[2:][::-1]
except:
pass
img = ImageData(img, max_size=max_size)
img.resize_image(mode='pad')
t1 = time.time()
logging.debug(f'Preparing image took: {t1 - t0}')
t0 = time.time()
boxes, probs, landmarks, mask_probs = self.det_model.detect(img.transformed_image, threshold=threshold)
t1 = time.time()
logging.debug(f'Detection took: {t1 - t0}')
ret = []
if not isinstance(boxes, type(None)):
t0 = time.time()
for i in range(len(boxes)):
# Translate points to original image size
bbox = self.reproject_points(boxes[i], img.scale_factor)
landmark = self.reproject_points(landmarks[i], img.scale_factor)
det_score = probs[i]
if not isinstance(mask_probs, type(None)):
mask_prob = mask_probs[i]
else:
mask_prob = None
# Crop faces from original image instead of resized to improve quality
_crop = face_align.norm_crop(img.orig_image, landmark=landmark)
# _crop = cv2.resize(_crop, (48, 48))
# _crop = cv2.resize(_crop, (112, 112))
embedding = self.rec_model.get_embedding(_crop)
embedding_norm = norm(embedding)
normed_embedding = embedding / embedding_norm
# print(normed_embedding.shape)
# print(_crop.shape)
facedata = None
gender = None
age = None
if return_face_data:
facedata = _crop
# if self.ga_model:
# if extract_ga:
# gender, age = self.ga_model.get(_crop)
# gender = int(gender)
face = Face(bbox=bbox, landmark=landmark, det_score=det_score, facedata=facedata, num_det=i, scale=img.scale_factor, mask_prob=mask_prob, normed_embedding=normed_embedding)
ret.append(face)
t1 = time.time()
logging.debug(f'Cropping {len(boxes)} faces took: {t1 - t0}')
# Embedding detected faces
# if extract_embedding and self.rec_model:
# t0 = time.time()
# ret = [e for e in self.embed_faces(ret)]
# t1 = time.time()
# logging.debug(f'Embedding {len(boxes)} faces took: {t1 - t0}')
tf = time.time()
logging.debug(f'Full processing took: {tf - ts}')
return ret