def detect(self, img, bbox):
crop_image, detail = self.crop_image(img, bbox)
crop_image = (crop_image - 127.0) / 127.0
crop_image = np.array([np.transpose(crop_image, (2, 0, 1))]).astype(np.float32)
start = time.time()
raw = self.sess.run(None, {self.input_name: crop_image})[0][0]
end = time.time()
print("ONNX Inference Time: {:.6f}".format(end - start))
landmark = raw[0:136].reshape((-1, 2))
landmark[:, 0] = landmark[:, 0] * detail[1] + detail[3]
landmark[:, 1] = landmark[:, 1] * detail[0] + detail[2]
landmark = self.tracker.track(img, landmark)
_, PRY_3d = get_head_pose(landmark, img)
return landmark, PRY_3d[:, 0]
def detect(self, img, bbox):
crop_image, detail = self.crop_image(img, bbox)
crop_image = (crop_image - 127.0) / 127.0
crop_image = np.array([np.transpose(crop_image, (2, 0, 1))])
crop_image = torch.tensor(crop_image).float().cuda()
with torch.no_grad():
start = time.time()
raw = self.model(crop_image)[0].cpu().numpy()
end = time.time()
print("PyTorch Inference Time: {:.6f}".format(end - start))
landmark = raw[0:136].reshape((-1, 2))
landmark[:, 0] = landmark[:, 0] * detail[1] + detail[3]
landmark[:, 1] = landmark[:, 1] * detail[0] + detail[2]
landmark = self.tracker.track(img, landmark)
_, PRY_3d = get_head_pose(landmark, img)
return landmark, PRY_3d[:, 0]
def detect(self, img, bbox):
crop_image, detail = self.crop_image(img, bbox)
crop_image = (crop_image - 127.0) / 127.0
crop_image = np.array([np.transpose(crop_image, (2, 0, 1))]).astype(np.float32)
tmp_input = MNN.Tensor((1, 3, *self.detection_size), MNN.Halide_Type_Float, crop_image,
MNN.Tensor_DimensionType_Caffe)
self.input_tensor.copyFrom(tmp_input)
start = time.time()
self.interpreter.runSession(self.session)
raw = np.array(self.interpreter.getSessionOutput(self.session).getData())
end = time.time()
print("MNN Inference Time: {:.6f}".format(end - start))
landmark = raw[0:136].reshape((-1, 2))
landmark[:, 0] = landmark[:, 0] * detail[1] + detail[3]
landmark[:, 1] = landmark[:, 1] * detail[0] + detail[2]
landmark = self.tracker.track(img, landmark)
_, PRY_3d = get_head_pose(landmark, img)
return landmark, PRY_3d[:, 0]
def __getitem__(self, item):
"""Data augmentation function."""
dp = self.lst[item]
fname = dp['image_path']
keypoints = dp['keypoints']
bbox = dp['bbox']
if keypoints is not None:
if ".jpg" in fname:
image = jpeg.imread(fname)
# image = cv2.imread(fname)
else:
image = cv2.imread(fname)
label = np.array(keypoints, dtype=np.float).reshape((-1, 2))
bbox = np.array(bbox)
crop_image, label = self.augmentationCropImage(
image, bbox, label, self.training_flag)
if self.training_flag:
if random.uniform(0, 1) > 0.5:
crop_image, label = Mirror(crop_image,
label=label,
symmetry=symmetry)
if random.uniform(0, 1) > 0.0:
angle = random.uniform(-45, 45)
crop_image, label = Rotate_aug(crop_image,
label=label,
angle=angle)
if random.uniform(0, 1) > 0.5:
strength = random.uniform(0, 50)
crop_image, label = Affine_aug(crop_image,
strength=strength,
label=label)
if random.uniform(0, 1) > 0.5:
crop_image = self.color_augmentor(crop_image)
if random.uniform(0, 1) > 0.5:
crop_image = pixel_jitter(crop_image, 15)
if random.uniform(0, 1) > 0.5:
crop_image = Img_dropout(crop_image, 0.2)
if random.uniform(0, 1) > 0.5:
crop_image = Padding_aug(crop_image, 0.3)
reprojectdst, euler_angle = get_head_pose(label, crop_image)
PRY = euler_angle.reshape([-1]).astype(np.float32) / 90.
cla_label = np.zeros([4])
if dp['left_eye_close']:
cla_label[0] = 1
if dp['right_eye_close']:
cla_label[1] = 1
if dp['mouth_close']:
cla_label[2] = 1
if dp['big_mouth_open']:
cla_label[3] = 1
crop_image_height, crop_image_width, _ = crop_image.shape
# for point in label:
# crop_image = cv2.circle(crop_image, tuple(point.astype(np.int)), 3, (255, 0, 0), -1, 1)
# cv2.imshow("", crop_image)
# cv2.waitKey()
label = label.astype(np.float32)
label[:, 0] = label[:, 0] / crop_image_width
label[:, 1] = label[:, 1] / crop_image_height
crop_image = crop_image.astype(np.float32)
label = label.reshape([-1]).astype(np.float32)
cla_label = cla_label.astype(np.float32)
label = np.concatenate([label, PRY, cla_label], axis=0)
crop_image = (crop_image - 127.0) / 127.0
crop_image = np.transpose(crop_image, (2, 0, 1)).astype(np.float32)
return crop_image, label