def camera_demo():
args = build_argparser().parse_args()
model_xml = "openvino/480x640/FP32/dbface.xml" #<--- CPU
#model_xml = "openvino/480x640/FP16/dbface.xml" #<--- MYRIAD
model_bin = os.path.splitext(model_xml)[0] + ".bin"
ie = IECore()
net = ie.read_network(model=model_xml, weights=model_bin)
input_blob = next(iter(net.inputs))
exec_net = ie.load_network(network=net, device_name='CPU')
cap = cv2.VideoCapture(0)
cap.set(cv2.CAP_PROP_FRAME_WIDTH, 640)
cap.set(cv2.CAP_PROP_FRAME_HEIGHT, 480)
ok, frame = cap.read()
while ok:
img = frame[np.newaxis, :, :, :] # Batch size axis add
img = img.transpose((0, 3, 1, 2)) # NHWC to NCHW
objs = detect(exec_net, input_blob, img)
for obj in objs:
common.drawbbox(frame, obj)
cv2.imshow("demo DBFace", frame)
key = cv2.waitKey(1) & 0xFF
if key == ord('q'):
break
ok, frame = cap.read()
cap.release()
cv2.destroyAllWindows()
def camera_demo():
dbface = DBFace()
dbface.eval()
if HAS_CUDA:
dbface.cuda()
dbface.load("model/dbface.pth")
cap = cv2.VideoCapture(0)
cap.set(cv2.CAP_PROP_FRAME_WIDTH, 640)
cap.set(cv2.CAP_PROP_FRAME_HEIGHT, 480)
ok, frame = cap.read()
while ok:
objs = detect(dbface, frame)
for obj in objs:
common.drawbbox(frame, obj)
cv2.imshow("demo DBFace", frame)
key = cv2.waitKey(1) & 0xFF
if key == ord('q'):
break
ok, frame = cap.read()
cap.release()
cv2.destroyAllWindows()
def _draw_box(img: np.array, face_infos: List[Dict], output_filename):
for face_info in face_infos:
common.drawbbox(img, face_info)
# common.imwrite("detect_result/" + common.file_name_no_suffix(filename) + ".draw.jpg", image)
if output_filename != "":
common.imwrite(output_filename, img)
def camera_demo():
interpreter = Interpreter(model_path='model_float32.tflite', num_threads=4)
interpreter.allocate_tensors()
input_blob = interpreter.get_input_details()
output_blob = interpreter.get_output_details()
cap = cv2.VideoCapture(0)
cap.set(cv2.CAP_PROP_FRAME_WIDTH, 640)
cap.set(cv2.CAP_PROP_FRAME_HEIGHT, 480)
while True:
ok, frame = cap.read()
if not ok:
continue
img = cv2.resize(frame, (640, 480))
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
img = img.astype(np.float32)
img = img[np.newaxis, :, :, :]
objs = detect(interpreter, input_blob, output_blob, img)
for obj in objs:
common.drawbbox(frame, obj)
cv2.imshow("demo DBFace", frame)
key = cv2.waitKey(1) & 0xFF
if key == ord('q'):
break
cap.release()
cv2.destroyAllWindows()
def detect_image(model, file):
image = common.imread(file)
objs = detect(model, image)
for obj in objs:
common.drawbbox(image, obj)
common.imwrite("detect_result/" + common.file_name_no_suffix(file) + ".draw.jpg", image)
def detect_image(model, file):
image = common.imread(file)
objs = detect(model, image)
for obj in objs:
# 增加对五官的检测
frame = organ_detection(image, obj)
common.drawbbox(image, obj)
cv2.imshow('demo', image)
cv2.waitKey()
def train_epoch(self, epoch):
for indbatch, (images, heatmap_gt, heatmap_posweight, reg_tlrb,
reg_mask, landmark_gt, landmark_mask, num_objs,
keep_mask) in enumerate(self.train_loader):
self.iter += 1
batch_objs = sum(num_objs)
batch_size = self.batch_size
if batch_objs == 0:
batch_objs = 1
heatmap_gt = heatmap_gt.to(self.gpu_master)
heatmap_posweight = heatmap_posweight.to(self.gpu_master)
keep_mask = keep_mask.to(self.gpu_master)
reg_tlrb = reg_tlrb.to(self.gpu_master)
reg_mask = reg_mask.to(self.gpu_master)
landmark_gt = landmark_gt.to(self.gpu_master)
landmark_mask = landmark_mask.to(self.gpu_master)
images = images.to(self.gpu_master)
hm, tlrb, landmark = self.model(images)
hm = hm.sigmoid()
hm = torch.clamp(hm, min=1e-4, max=1 - 1e-4)
tlrb = torch.exp(tlrb)
hm_loss = self.focal_loss(
hm, heatmap_gt, heatmap_posweight,
keep_mask=keep_mask) / batch_objs
reg_loss = self.giou_loss(tlrb, reg_tlrb, reg_mask) * 5
landmark_loss = self.landmark_loss(landmark, landmark_gt,
landmark_mask) * 0.1
loss = hm_loss + reg_loss + landmark_loss
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
epoch_flt = epoch + indbatch / self.per_epoch_batchs
if indbatch % 10 == 0:
log.info(
f"iter: {self.iter}, lr: {self.lr:g}, epoch: {epoch_flt:.2f}, loss: {loss.item():.2f}, hm_loss: {hm_loss.item():.2f}, "
f"box_loss: {reg_loss.item():.2f}, lmdk_loss: {landmark_loss.item():.5f}"
)
if indbatch % 1000 == 0:
log.info("save hm")
hm_image = hm[0, 0].cpu().data.numpy()
common.imwrite(f"{jobdir}/imgs/hm_image.jpg", hm_image * 255)
common.imwrite(f"{jobdir}/imgs/hm_image_gt.jpg",
heatmap_gt[0, 0].cpu().data.numpy() * 255)
image = np.clip(
(images[0].permute(1, 2, 0).cpu().data.numpy() * self.std +
self.mean) * 255, 0, 255).astype(np.uint8)
outobjs = eval_tool.detect_images_giou_with_netout(
hm, tlrb, landmark, threshold=0.1, ibatch=0)
im1 = image.copy()
for obj in outobjs:
common.drawbbox(im1, obj)
common.imwrite(f"{jobdir}/imgs/train_result.jpg", im1)
if flags[index] != 0:
continue
keep.append(obj)
for j in range(index + 1, len(objs)):
if flags[j] == 0 and obj.iou(objs[j]) > iou:
flags[j] = 1
return keep
mean = [0.408, 0.447, 0.47]
std = [0.289, 0.274, 0.278]
trial_name = "small-H-dense-wide64-UCBA"
jobdir = f"jobs/{trial_name}"
image = common.imread("imgs/selfie.jpg")
model = DBFace(has_landmark=True, wide=64, has_ext=True, upmode="UCBA")
#model.load(f"{jobdir}/models/150.pth")
model.load_from_zoo()
model.eval()
model.cuda()
outs = eval_tool.detect_image(model, image, mean, std, 0.3)
outs = nms(outs, 0.2)
print("objs = %d" % len(outs))
for obj in outs:
common.drawbbox(image, obj)
common.imwrite("test_result/test.jpg", image)
print("ok")
