def get_face(self, srcimg, threshold=0.5):
height, width = srcimg.shape[:2]
self.img_h_new, self.img_w_new, self.scale_h, self.scale_w = self.transform(
height, width)
dets, lms = self.inference_opencv(srcimg, threshold)
boxs, face_rois = [], []
for i in range(dets.shape[0]):
# boxes, score = dets[i, :4], dets[i, 4]
box = [
int(dets[i, 0]),
int(dets[i, 1]),
int(dets[i, 2]),
int(dets[i, 3])
]
face_roi = srcimg[box[1]:box[3], box[0]:box[2]]
if len(lms) > 0:
lm = lms[i, :]
if self.align:
face_roi = align_process(srcimg, np.array(box),
np.array(lm).reshape(-1, 2),
(224, 224))
box.extend(list(map(int, lm.tolist())))
boxs.append(tuple(box))
face_rois.append(face_roi)
return boxs, face_rois
def get_face(self, img):
h, w, _ = img.shape
blob = cv2.dnn.blobFromImage(img, size=self.input_shape)
self.net.setInput(blob)
# output_names = ['loc', 'conf']
# loc, conf = self.net.forward(output_names)
loc, conf = self.net.forward(self.net.getUnconnectedOutLayersNames())
# Decode bboxes and landmarks
pb = PriorBox(input_shape=self.input_shape, output_shape=(w, h))
dets = pb.decode(np.squeeze(loc, axis=0), np.squeeze(conf, axis=0))
# Ignore low scores
idx = np.where(dets[:, -1] > self.conf_thresh)[0]
dets = dets[idx]
# NMS
if dets.shape[0] > 0:
dets = nms(dets, self.nms_thresh)
faces = dets[:self.keep_top_k, :]
# Draw boudning boxes and landmarks on the original image
boxs, face_rois = [], []
for i in range(faces.shape[0]):
# score = faces[i,-1]
box = (faces[i, :4]).astype(np.int32).tolist()
face_roi = img[box[1]:box[3], box[0]:box[2]]
landmark = faces[i, 4:14].reshape((5, 2))
if self.align:
face_roi = align_process(img, faces[i, :4], landmark,
(224, 224))
box.extend(landmark.astype(np.int32).ravel().tolist())
boxs.append(tuple(box))
face_rois.append(face_roi)
return boxs, face_rois
else:
return [], []
def detect(self, srcimg, threshold=0.5):
height, width = srcimg.shape[:2]
self.img_h_new, self.img_w_new, self.scale_h, self.scale_w = self.transform(
height, width)
dets, lms = self.inference_opencv(srcimg, threshold)
drawimg, face_rois = srcimg.copy(), []
for i in range(dets.shape[0]):
boxes, score = dets[i, :4], dets[i, 4]
cv2.rectangle(drawimg, (int(boxes[0]), int(boxes[1])),
(int(boxes[2]), int(boxes[3])), (0, 0, 255),
thickness=2)
face_roi = srcimg[int(boxes[1]):int(boxes[3]),
int(boxes[0]):int(boxes[2])]
if len(lms) > 0:
lm = lms[
i, :] ###landmarks: numpy array, n x 10 (x1, y1 ... x5,y5)
for j in range(0, 5):
cv2.circle(drawimg, (int(lm[j * 2]), int(lm[j * 2 + 1])),
2, (0, 255, 0),
thickness=-1)
# cv2.putText(drawimg, str(j), (int(lm[j * 2]), int(lm[j * 2 + 1]) + 12), cv2.FONT_HERSHEY_DUPLEX, 1,(0, 0, 255))
if self.align:
face_roi = align_process(srcimg, np.array(boxes),
np.array(lm).reshape(-1, 2),
(224, 224))
face_rois.append(face_roi)
return drawimg, face_rois
def detect(self, srcimg):
bounding_boxes, landmarks = self.mtcnn.detect(
srcimg
) ###landmarks: numpy array, n x 10 (x1, x2 ... x5, y1, y2 ..y5)
if len(bounding_boxes) == 0:
return srcimg, []
drawimg, face_rois = srcimg.copy(), []
for i in range(bounding_boxes.shape[0]):
# score = bounding_boxes[i,4]
x1, y1, x2, y2 = (bounding_boxes[i, :4]).astype(np.int32)
cv2.rectangle(drawimg, (x1, y1), (x2, y2), (0, 0, 255),
thickness=2)
face_roi = srcimg[y1:y2, x1:x2]
landmark = landmarks[i, :].reshape((2, 5)).T
if self.align:
face_roi = align_process(srcimg, bounding_boxes[i, :4],
landmark, (224, 224))
landmark = landmark.astype(np.int32)
for j in range(5):
cv2.circle(drawimg, (landmark[j, 0], landmark[j, 1]),
2, (0, 255, 0),
thickness=-1)
# cv2.putText(drawimg, str(j), (landmark[j, 0], landmark[j, 1] + 12), cv2.FONT_HERSHEY_DUPLEX, 1, (0, 0, 255))
face_rois.append(face_roi)
return drawimg, face_rois
def get_face(self, srcimg):
ori_h, ori_w, _ = srcimg.shape
LONG_SIDE = self.long_side
if self.long_side == -1:
max_size = max(ori_w, ori_h)
LONG_SIDE = max(32, max_size - max_size % 32)
if ori_h > ori_w:
scale_h = LONG_SIDE / ori_h
tar_w = int(ori_w * scale_h)
tar_w = tar_w - tar_w % 32
tar_w = max(32, tar_w)
tar_h = LONG_SIDE
else:
scale_w = LONG_SIDE / ori_w
tar_h = int(ori_h * scale_w)
tar_h = tar_h - tar_h % 32
tar_h = max(32, tar_h)
tar_w = LONG_SIDE
scale_w = tar_w * 1.0 / ori_w
scale_h = tar_h * 1.0 / ori_h
image = cv2.resize(srcimg, (tar_w, tar_h))
image = image[..., ::-1]
image = image.astype(np.float64)
# image = (image - hyp['mean']) / hyp['std']
image /= 255.0
image = np.transpose(image, [2, 0, 1])
image = np.expand_dims(image, axis=0)
with torch.no_grad():
image = torch.from_numpy(image)
image = image.to(self.device).float()
pred = self.net(image)[0]
pred = non_max_suppression(pred, 0.25, 0.35, multi_label=False, classes=0, agnostic=False, land=True,
point_num=self.point_num)
try:
det = pred[0].cpu().detach().numpy()
srcimg = srcimg.astype(np.uint8)
det[:, :4] = det[:, :4] / np.array([scale_w, scale_h] * 2)
det[:, 5:5 + self.point_num * 2] = det[:, 5:5 + self.point_num * 2] / np.array([scale_w, scale_h] * self.point_num)
except:
det = []
boxs, face_rois = [], []
for b in det:
b = list(map(int, b))
del b[4] ### delte score
boxs.append(b)
face_roi = srcimg[b[1]:b[3], b[0]:b[2]]
if self.align:
face_roi = align_process(srcimg, np.array(b[:4]), np.array(b[4:14]).reshape(-1, 2), (224, 224))
face_rois.append(face_roi)
return boxs, face_rois
def get_face(self, srcimg, threshold=0.3, nms_iou=0.3):
image = common.pad(srcimg).astype(np.float32)
image = ((image / 255.0 - self.mean) / self.std)
image = image.transpose(2, 0, 1)
with torch.no_grad():
torch_image = torch.from_numpy(image)[None]
torch_image = torch_image.to(self.device)
hm, box, landmark = self.net(torch_image)
hm_pool = F.max_pool2d(hm, 3, 1, 1)
scores, indices = ((hm == hm_pool).float() * hm).view(
1, -1).cpu().topk(1000)
hm_height, hm_width = hm.shape[2:]
scores = scores.squeeze()
indices = indices.squeeze()
ys = list((indices / hm_width).int().data.numpy())
xs = list((indices % hm_width).int().data.numpy())
scores = list(scores.data.numpy())
box = box.cpu().squeeze().data.numpy()
landmark = landmark.cpu().squeeze().data.numpy()
stride = 4
objs = []
for cx, cy, score in zip(xs, ys, scores):
if score < threshold:
break
x, y, r, b = box[:, cy, cx]
xyrb = (np.array([cx, cy, cx, cy]) + [-x, -y, r, b]) * stride
x5y5 = landmark[:, cy, cx]
x5y5 = (common.exp(x5y5 * 4) + ([cx] * 5 + [cy] * 5)) * stride
box_landmark = list(zip(x5y5[:5], x5y5[5:]))
objs.append(
common.BBox(0,
xyrb=xyrb,
score=score,
landmark=box_landmark))
objs = self.nms(objs, iou=nms_iou)
boxs, face_rois = [], []
for i, obj in enumerate(objs):
box, score, landmark = list(map(int,
obj.box)), obj.score, obj.landmark
landmark = [int(x) for t in obj.landmark for x in t]
boxs.append(box + landmark)
face_roi = srcimg[box[1]:box[3], box[0]:box[2]]
if self.align:
face_roi = align_process(srcimg, np.array(box),
np.array(landmark).reshape(-1, 2),
(224, 224))
face_rois.append(face_roi)
return boxs, face_rois
def get_face(self, srcimg):
bounding_boxes, landmarks = self.retinaface.detect_faces(srcimg)
boxs, face_rois = [], []
for i in range(bounding_boxes.shape[0]):
# score = bounding_boxes[i,4]
box = (bounding_boxes[i, :4]).astype(np.int32).tolist()
face_roi = srcimg[box[1]:box[3], box[0]:box[2]]
landmark = landmarks[i, :].reshape((2, 5)).T
if self.align:
face_roi = align_process(srcimg, bounding_boxes[i, :4],
landmark, (224, 224))
box.extend(landmark.astype(np.int32).ravel().tolist())
boxs.append(tuple(box))
face_rois.append(face_roi)
return boxs, face_rois
def detect(self, img):
h, w, _ = img.shape
blob = cv2.dnn.blobFromImage(img, size=self.input_shape)
self.net.setInput(blob)
# output_names = ['loc', 'conf']
# loc, conf = self.net.forward(output_names)
loc, conf = self.net.forward(self.net.getUnconnectedOutLayersNames())
# Decode bboxes and landmarks
pb = PriorBox(input_shape=self.input_shape, output_shape=(w, h))
dets = pb.decode(np.squeeze(loc, axis=0), np.squeeze(conf, axis=0))
# Ignore low scores
idx = np.where(dets[:, -1] > self.conf_thresh)[0]
dets = dets[idx]
# NMS
if dets.shape[0] > 0:
dets = nms(dets, self.nms_thresh)
faces = dets[:self.keep_top_k, :]
# Draw boudning boxes and landmarks on the original image
drawimg, face_rois = img.copy(), []
for i in range(faces.shape[0]):
# score = faces[i,-1]
x1, y1, x2, y2 = (faces[i, :4]).astype(np.int32)
cv2.rectangle(drawimg, (x1, y1), (x2, y2), (0, 0, 255),
thickness=2)
face_roi = img[y1:y2, x1:x2]
landmark = faces[i, 4:14].reshape((5, 2))
if self.align:
face_roi = align_process(img, faces[i, :4], landmark,
(224, 224))
landmark = landmark.astype(np.int32)
for j in range(5):
cv2.circle(drawimg, (landmark[j, 0], landmark[j, 1]),
2, (0, 255, 0),
thickness=-1)
# cv2.putText(drawimg, str(j), (landmark[j, 0], landmark[j, 1] + 12), cv2.FONT_HERSHEY_DUPLEX, 1, (0, 0, 255))
face_rois.append(face_roi)
return drawimg, face_rois
else:
print('No faces found.')
return img, []
def detect(self, srcimg):
ori_h, ori_w, _ = srcimg.shape
LONG_SIDE = self.long_side
if self.long_side == -1:
max_size = max(ori_w, ori_h)
LONG_SIDE = max(32, max_size - max_size % 32)
if ori_h > ori_w:
scale_h = LONG_SIDE / ori_h
tar_w = int(ori_w * scale_h)
tar_w = tar_w - tar_w % 32
tar_w = max(32, tar_w)
tar_h = LONG_SIDE
else:
scale_w = LONG_SIDE / ori_w
tar_h = int(ori_h * scale_w)
tar_h = tar_h - tar_h % 32
tar_h = max(32, tar_h)
tar_w = LONG_SIDE
scale_w = tar_w * 1.0 / ori_w
scale_h = tar_h * 1.0 / ori_h
image = cv2.resize(srcimg, (tar_w, tar_h))
image = image[..., ::-1]
image = image.astype(np.float64)
# image = (image - hyp['mean']) / hyp['std']
image /= 255.0
image = np.transpose(image, [2, 0, 1])
image = np.expand_dims(image, axis=0)
with torch.no_grad():
image = torch.from_numpy(image)
image = image.to(self.device).float()
pred = self.net(image)[0]
pred = non_max_suppression(pred, 0.25, 0.35, multi_label=False, classes=0, agnostic=False, land=True,
point_num=self.point_num)
try:
det = pred[0].cpu().detach().numpy()
srcimg = srcimg.astype(np.uint8)
det[:, :4] = det[:, :4] / np.array([scale_w, scale_h] * 2)
det[:, 5:5 + self.point_num * 2] = det[:, 5:5 + self.point_num * 2] / np.array([scale_w, scale_h] * self.point_num)
except:
det = []
drawimg, face_rois = srcimg.copy(), []
for b in det:
# text = "{:.4f}".format(b[4])
b = list(map(int, b)) ###landmarks: numpy array, n x 10 (x1, y1 ... x5,y5)
cv2.rectangle(drawimg, (b[0], b[1]), (b[2], b[3]), (0, 0, 255), thickness=2)
cx, cy = b[0], b[1] + 12
# cv2.putText(drawimg, text, (cx, cy), cv2.FONT_HERSHEY_DUPLEX, 0.5, (255, 255, 255))
cv2.circle(drawimg, (b[5], b[6]), 1, (0, 0, 255), 4)
cv2.circle(drawimg, (b[7], b[8]), 1, (0, 255, 255), 4)
cv2.circle(drawimg, (b[9], b[10]), 1, (255, 0, 255), 4)
cv2.circle(drawimg, (b[11], b[12]), 1, (0, 255, 0), 4)
cv2.circle(drawimg, (b[13], b[14]), 1, (255, 0, 0), 4)
# for i in range(5):
# cv2.putText(drawimg, str(i), (b[2*i+5], b[2*i+6]+12), cv2.FONT_HERSHEY_DUPLEX, 1, (0, 0, 255))
face_roi = srcimg[b[1]:b[3], b[0]:b[2]]
if self.align:
face_roi = align_process(srcimg, np.array(b[:4]), np.array(b[5:15]).reshape(-1, 2), (224,224))
face_rois.append(face_roi)
return drawimg, face_rois
def detect(self, srcimg, threshold=0.3, nms_iou=0.3):
image = common.pad(srcimg).astype(np.float32)
image = ((image / 255.0 - self.mean) / self.std)
image = image.transpose(2, 0, 1)
with torch.no_grad():
torch_image = torch.from_numpy(image)[None]
torch_image = torch_image.to(self.device)
hm, box, landmark = self.net(torch_image)
hm_pool = F.max_pool2d(hm, 3, 1, 1)
scores, indices = ((hm == hm_pool).float() * hm).view(
1, -1).cpu().topk(1000)
hm_height, hm_width = hm.shape[2:]
scores = scores.squeeze()
indices = indices.squeeze()
ys = list((indices / hm_width).int().data.numpy())
xs = list((indices % hm_width).int().data.numpy())
scores = list(scores.data.numpy())
box = box.cpu().squeeze().data.numpy()
landmark = landmark.cpu().squeeze().data.numpy()
stride = 4
objs = []
for cx, cy, score in zip(xs, ys, scores):
if score < threshold:
break
x, y, r, b = box[:, cy, cx]
xyrb = (np.array([cx, cy, cx, cy]) + [-x, -y, r, b]) * stride
x5y5 = landmark[:, cy, cx]
x5y5 = (common.exp(x5y5 * 4) + ([cx] * 5 + [cy] * 5)) * stride
box_landmark = list(zip(x5y5[:5], x5y5[5:]))
objs.append(
common.BBox(0,
xyrb=xyrb,
score=score,
landmark=box_landmark))
objs = self.nms(objs, iou=nms_iou)
drawimg, face_rois = srcimg.copy(), []
for i, obj in enumerate(objs):
box, score, landmark = list(map(int,
obj.box)), obj.score, obj.landmark
landmark = [int(x) for t in obj.landmark for x in t]
# landmark = sum(obj.landmark, ())
# landmark = list(itertools.chain.from_iterable(obj.landmark))
cv2.rectangle(drawimg, (box[0], box[1]), (box[2], box[3]),
(0, 0, 255),
thickness=2)
for j in range(0, 5):
cv2.circle(drawimg, (landmark[j * 2], landmark[j * 2 + 1]),
2, (0, 255, 0),
thickness=-1)
# cv2.putText(drawimg, str(j), (landmark[j * 2], landmark[j * 2 + 1] + 12), cv2.FONT_HERSHEY_DUPLEX, 1, (0, 0, 255))
face_roi = srcimg[box[1]:box[3], box[0]:box[2]]
# import imutils
# angle = np.rad2deg(np.arctan2(landmark[3] - landmark[1], landmark[2] - landmark[0]))
# if angle != 0 and self.align:
# face_roi = imutils.rotate(face_roi, angle)
if self.align:
face_roi = align_process(srcimg, np.array(box),
np.array(landmark).reshape(-1, 2),
(224, 224))
face_rois.append(face_roi)
return drawimg, face_rois
