def infer_fast(net, img, net_input_height_size, stride, upsample_ratio, cpu,
pad_value=(0, 0, 0), img_mean=(128, 128, 128), img_scale=1/256):
height, width, _ = img.shape
scale = net_input_height_size / height
scaled_img = cv2.resize(img, (0, 0), fx=scale, fy=scale, interpolation=cv2.INTER_CUBIC)
scaled_img = normalize(scaled_img, img_mean, img_scale)
min_dims = [net_input_height_size, max(scaled_img.shape[1], net_input_height_size)]
padded_img, pad = pad_width(scaled_img, stride, pad_value, min_dims)
tensor_img = torch.from_numpy(padded_img).permute(2, 0, 1).unsqueeze(0).float()
if not cpu:
tensor_img = tensor_img.cuda()
stages_output = net(tensor_img)
stage2_heatmaps = stages_output[-2]
heatmaps = np.transpose(stage2_heatmaps.squeeze().cpu().data.numpy(), (1, 2, 0))
heatmaps = cv2.resize(heatmaps, (0, 0), fx=upsample_ratio, fy=upsample_ratio, interpolation=cv2.INTER_CUBIC)
stage2_pafs = stages_output[-1]
pafs = np.transpose(stage2_pafs.squeeze().cpu().data.numpy(), (1, 2, 0))
pafs = cv2.resize(pafs, (0, 0), fx=upsample_ratio, fy=upsample_ratio, interpolation=cv2.INTER_CUBIC)
return heatmaps, pafs, scale, pad
def infer_fast(net, img, net_input_height_size, stride, upsample_ratio, cpu,
pad_value=(0, 0, 0), img_mean=(128, 128, 128), img_scale=1/256, openvino=False):
height, width, _ = img.shape
scale = net_input_height_size / height
scaled_img = cv2.resize(img, (0, 0), fx=scale, fy=scale, interpolation=cv2.INTER_CUBIC)
if not openvino: # OpenVINO model doesn't require input to be normalized
scaled_img = normalize(scaled_img, img_mean, img_scale)
min_dims = [net_input_height_size, max(scaled_img.shape[1], net_input_height_size)]
padded_img, pad = pad_width(scaled_img, stride, pad_value, min_dims)
if openvino:
reshaped_img = np.transpose(padded_img, (2, 0, 1))
res = net.infer(inputs={next(iter(net.inputs)): reshaped_img})
pafs = res['Mconv7_stage2_L1']
heatmaps = res['Mconv7_stage2_L2']
pafs = np.transpose(pafs.squeeze(), (1, 2, 0))
heatmaps = np.transpose(heatmaps.squeeze(), (1, 2, 0))
else:
tensor_img = torch.from_numpy(padded_img).permute(2, 0, 1).unsqueeze(0).float()
if not cpu:
tensor_img = tensor_img.cuda()
stages_output = net(tensor_img)
stage2_heatmaps = stages_output[-2]
heatmaps = np.transpose(stage2_heatmaps.squeeze().cpu().data.numpy(), (1, 2, 0))
stage2_pafs = stages_output[-1]
pafs = np.transpose(stage2_pafs.squeeze().cpu().data.numpy(), (1, 2, 0))
heatmaps = cv2.resize(heatmaps, (0, 0), fx=upsample_ratio, fy=upsample_ratio, interpolation=cv2.INTER_CUBIC)
pafs = cv2.resize(pafs, (0, 0), fx=upsample_ratio, fy=upsample_ratio, interpolation=cv2.INTER_CUBIC)
return heatmaps, pafs, scale, pad
def infer_fast(net,
img,
net_input_height_size,
stride,
upsample_ratio,
cpu,
pad_value=(0, 0, 0),
img_mean=(128, 128, 128),
img_scale=1 / 256):
height, width, _ = img.shape
scale = net_input_height_size / height
scaled_img = cv2.resize(img, (0, 0),
fx=scale,
fy=scale,
interpolation=cv2.INTER_CUBIC)
scaled_img = normalize(scaled_img, img_mean, img_scale)
min_dims = [
net_input_height_size,
max(scaled_img.shape[1], net_input_height_size)
]
padded_img, pad = pad_width(scaled_img, stride, pad_value, min_dims)
tensor_img = torch.from_numpy(padded_img).permute(2, 0,
1).unsqueeze(0).float()
if not cpu:
tensor_img = tensor_img.cuda()
stages_output = net(tensor_img)
stage2_heatmaps = stages_output[-2]
heatmaps = np.transpose(stage2_heatmaps.squeeze().cpu().data.numpy(),
(1, 2, 0))
heatmaps = cv2.resize(heatmaps, (0, 0),
fx=upsample_ratio,
fy=upsample_ratio,
interpolation=cv2.INTER_CUBIC)
# for i in range(heatmaps.shape[-1]):
# heatmap = heatmaps[:,:,i]
# heat_map = cv2.resize(heatmap, (img.shape[1], img.shape[0]))
# heat_map = np.fmin(np.fmax(heat_map, 0.0), 1.0)
# heat_map = np.expand_dims(np.uint8(heat_map * 255), -1)
# debug_map = img // 2 + heat_map // 2
# win_name = 'keypoint: {}'.format(kp_names[i])
# cv2.imshow(win_name, debug_map)
# cv2.moveWindow(win_name, 0, 0)
# cv2.waitKey(0)
# cv2.destroyWindow(win_name)
stage2_pafs = stages_output[-1]
pafs = np.transpose(stage2_pafs.squeeze().cpu().data.numpy(), (1, 2, 0))
pafs = cv2.resize(pafs, (0, 0),
fx=upsample_ratio,
fy=upsample_ratio,
interpolation=cv2.INTER_CUBIC)
# for i in range(pafs.shape[-1] // 2):
# paf = np.zeros((pafs.shape[0], pafs.shape[1], 3), dtype=np.float32)
# pair = BODY_PARTS_PAF_IDS[i]
# paf[:,:,2] = np.abs(pafs[:,:,pair[0]])
# paf[:,:,1] = np.abs(pafs[:,:,pair[1]])
# paf = np.fmin(np.fmax(paf, -1.0), 1.0)
# paf = cv2.resize(paf, (img.shape[1], img.shape[0]))
# paf = np.uint8(abs(paf) * 255)
# debug_map = img // 2 + paf // 2
# win_name = 'part: {}'.format(pf_names[i])
# cv2.imshow(win_name, debug_map)
# cv2.moveWindow(win_name, 0, 0)
# cv2.waitKey(0)
# cv2.destroyWindow(win_name)
return heatmaps, pafs, scale, pad
def infer_fast(model,
img,
net_input_height_size,
stride,
upsample_ratio,
cpu,
pad_value=(0, 0, 0),
img_mean=(128, 128, 128),
img_scale=1 / 256):
"""[summary]
Args:
model ([type]): [description]
img ([type]): [description]
net_input_height_size ([type]): [description]
stride ([type]): [description]
upsample_ratio ([type]): [description]
cpu ([type]): [description]
pad_value (tuple, optional): [description]. Defaults to (0, 0, 0).
img_mean (tuple, optional): [description]. Defaults to (128, 128, 128).
img_scale ([type], optional): [description]. Defaults to 1/256.
Returns:
[type]: [description]
"""
height, _, _ = img.shape
scale = net_input_height_size / height
scaled_img = cv2.resize(img, (0, 0),
fx=scale,
fy=scale,
interpolation=cv2.INTER_CUBIC)
scaled_img = normalize(scaled_img, img_mean, img_scale)
min_dims = [
net_input_height_size,
max(scaled_img.shape[1], net_input_height_size)
]
padded_img, pad = pad_width(scaled_img, stride, pad_value, min_dims)
tensor_img = torch.from_numpy(padded_img).permute(2, 0,
1).unsqueeze(0).float()
if not cpu:
tensor_img = tensor_img.cuda()
stages_output = model(tensor_img)
stage2_heatmaps = stages_output[-2]
heatmaps = np.transpose(stage2_heatmaps.squeeze().cpu().data.numpy(),
(1, 2, 0))
heatmaps = cv2.resize(heatmaps, (0, 0),
fx=upsample_ratio,
fy=upsample_ratio,
interpolation=cv2.INTER_CUBIC)
stage2_pafs = stages_output[-1]
pafs = np.transpose(stage2_pafs.squeeze().cpu().data.numpy(), (1, 2, 0))
pafs = cv2.resize(pafs, (0, 0),
fx=upsample_ratio,
fy=upsample_ratio,
interpolation=cv2.INTER_CUBIC)
return heatmaps, pafs, scale, pad