def visualize(img, seg_pred, exist_pred):
img = cv2.cvtColor(img, cv2.COLOR_RGB2BGR)
img = cv2.add(img, (-50, -50, -50, 0))
lane_img = np.zeros_like(img)
color = np.array([[255, 125, 0], [0, 255, 0], [0, 0, 255], [0, 255, 255]],
dtype='uint8') #b, g, r, y
coord_mask = np.argmax(seg_pred, axis=0)
exist = [1 if exist_pred[0, i] > 0.5 else 0 for i in range(4)]
lines = []
for i in getLane.prob2lines_CULane(seg_pred, exist):
print(i)
if len(i) < 8:
continue
i1_x = []
i1_y = []
for j in i:
i1_x.append(j[0])
i1_y.append(j[1])
lines.append(np.polyfit(np.array(i1_x), np.array(i1_y), 2))
yMax = img.shape[1]
for li in range(len(lines)):
#if exist_pred[0, li] > 0.5:
# lane_img[coord_mask == (li + 1)] = color[li]
for Y in np.linspace(0, yMax - 1, yMax):
X = lines[li][0] * Y**2 + lines[li][1] * Y + lines[li][2]
if (0 > Y) or (Y > 750):
continue
if (0 > X) or (X > 320):
continue
img[int(X), int(Y)] = color[li]
#img = cv2.addWeighted(src1=lane_img, alpha=0.8, src2=img, beta=1., gamma=0.)
return img
def post_processor(arg):
img_queue, arg_visualize = arg
while True:
if not img_queue.empty():
x, seg_pred, exist_pred = img_queue.get()
seg_pred = seg_pred.numpy()[0]
exist_pred = exist_pred.numpy()
exist = [1 if exist_pred[0, i] > 0.5 else 0 for i in range(4)]
print(exist)
for i in getLane.prob2lines_CULane(seg_pred, exist):
print(i)
if arg_visualize:
frame = x.squeeze().permute(1, 2, 0).numpy()
img = visualize(frame, seg_pred, exist_pred)
cv2.imshow('input_video', frame)
cv2.imshow("output_video", img)
if cv2.waitKey(1) & 0xFF == ord('q'):
break
else:
pass
def main():
args = parse_args()
img_path = args.img_path
weight_path = args.weight_path
img = cv2.imread(img_path)
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
img = transform_img({'img': img})['img']
x = transform_to_net({'img': img})['img']
x.unsqueeze_(0)
save_dict = torch.load(weight_path, map_location='cpu')
net.load_state_dict(save_dict['net'])
net.eval()
seg_pred, exist_pred = net(x)[:2]
seg_pred = seg_pred.detach().cpu().numpy()
exist_pred = exist_pred.detach().cpu().numpy()
seg_pred = seg_pred[0]
exist = [1 if exist_pred[0, i] > 0.5 else 0 for i in range(4)]
img = cv2.cvtColor(img, cv2.COLOR_RGB2BGR)
lane_img = np.zeros_like(img)
color = np.array([[255, 125, 0], [0, 255, 0], [0, 0, 255], [0, 255, 255]], dtype='uint8')
coord_mask = np.argmax(seg_pred, axis=0)
for i in range(0, 4):
if exist_pred[0, i] > 0.5:
lane_img[coord_mask == (i + 1)] = color[i]
img = cv2.addWeighted(src1=lane_img, alpha=0.8, src2=img, beta=1., gamma=0.)
cv2.imwrite("demo/demo_result.jpg", img)
print(" ")
print("Lane Lines:")
for x in getLane.prob2lines_CULane(seg_pred, exist):
print(x)
img = cv2.polylines(img, np.int32([np.array(x)]), 0, (0, 0, 255), thickness=3)
#
cv2.imwrite("demo/demo_result2.png", img)
print(" ")
print("exist: ", exist)
print("exist_pred: ", exist_pred)
if args.visualize:
print([1 if exist_pred[0, i] > 0.5 else 0 for i in range(4)])
cv2.imshow("", img)
cv2.waitKey(0)
cv2.destroyAllWindows()
with torch.no_grad():
for batch_idx, sample in enumerate(test_loader):
img = sample['img'].to(device)
img_name = sample['img_name']
seg_pred, exist_pred = net(img)[:2]
seg_pred = F.softmax(seg_pred, dim=1)
seg_pred = seg_pred.detach().cpu().numpy()
exist_pred = exist_pred.detach().cpu().numpy()
for b in range(len(seg_pred)):
seg = seg_pred[b]
exist = [1 if exist_pred[b, i] > 0.5 else 0 for i in range(4)]
lane_coords = getLane.prob2lines_CULane(seg,
exist,
resize_shape=(590, 1640),
y_px_gap=20,
pts=18)
path_tree = split_path(img_name[b])
save_dir, save_name = path_tree[-3:-1], path_tree[-1]
save_dir = os.path.join(out_path, *save_dir)
save_name = save_name[:-3] + "lines.txt"
save_name = os.path.join(save_dir, save_name)
if not os.path.exists(save_dir):
os.makedirs(save_dir)
with open(save_name, "w") as f:
for l in lane_coords:
for (x, y) in l:
print("{} {}".format(x, y), end=" ", file=f)
def main():
#
# determine which computer/platform we are running on
if (os.name == "posix"):
os_list = os.uname()
if (os_list[0] == "Darwin"):
pf_detected = 'MAC'
elif (os_list[0] == "Linux"):
if (os_list[1] == 'en4119351l'):
pf_detected = 'Quadro'
elif (os_list[1] == '19fef43c2174'):
pf_detected = 'Exxact'
elif (os_list[1] == 'EN4113948L'):
pf_detected = 'Kevin'
else:
pf_detected = 'PC'
# set the root path based on the computer/platform
# rootPath is path to directory in which webots/ and imdata/ directories reside
if (pf_detected == 'MAC'):
rootPath = '/Users/mes/Documents/ASU-Classes/Research/Ben-Amor/code/'
elif (pf_detected == 'Quadro'):
rootPath = '/home/local/ASUAD/mestric1/Documents/AVCES/'
elif (pf_detected == 'Exxact'):
rootPath = '/home/dockeruser/Documents/AVCES/'
elif (pf_detected == 'Kevin'):
rootPath = '/home/local/ASUAD/mestric1/Documents/AVCES/'
elif (pf_detected == 'PC'):
# rootPath = 'C:\Users\cesar\Desktop\Furi\'
print("PC platform detected. Exiting.")
exit()
else:
print("Computer/Platform not detected. Exiting.")
exit()
#
# args = parse_args()
# CCT007-Scene-005 has 153 frames
# CCT007-Scene-009 has 218 frames
# CCT007-Scene-045 has 665 frames
for fno in range(1, 666):
img_path = rootPath + "imdata/video/processed/CCT007/CCT007-Scene-045/Run/rgb{0:06d}.png".format(
fno)
weight_path = "experiments/exp10/exp10_best.pth"
img = cv2.imread(img_path)
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
img = transform_img({'img': img})['img']
x = transform_to_net({'img': img})['img']
x.unsqueeze_(0)
save_dict = torch.load(weight_path, map_location='cpu')
net.load_state_dict(save_dict['net'])
net.eval()
seg_pred, exist_pred = net(x)[:2]
seg_pred = seg_pred.detach().cpu().numpy()
exist_pred = exist_pred.detach().cpu().numpy()
seg_pred = seg_pred[0]
exist = [1 if exist_pred[0, i] > 0.5 else 0 for i in range(4)]
img = cv2.cvtColor(img, cv2.COLOR_RGB2BGR)
lane_img = np.zeros_like(img)
color = np.array(
[[255, 125, 0], [0, 255, 0], [0, 0, 255], [0, 255, 255]],
dtype='uint8')
coord_mask = np.argmax(seg_pred, axis=0)
for i in range(0, 4):
if exist_pred[0, i] > 0.5:
lane_img[coord_mask == (i + 1)] = color[i]
img = cv2.addWeighted(src1=lane_img,
alpha=0.8,
src2=img,
beta=1.,
gamma=0.)
# cv2.imwrite("demo/demo_result.jpg", img)
# print(" ")
# print("Lane Lines:")
for x in getLane.prob2lines_CULane(seg_pred, exist):
# print(x)
img = cv2.polylines(img,
np.int32([np.array(x)]),
0, (0, 0, 255),
thickness=3)
#
img_rsz = cv2.resize(img, (1280, 720),
interpolation=cv2.INTER_LANCZOS4)
cv2.imwrite(
rootPath +
"imdata/video/processed/CCT007/CCT007-Scene-045/Lane/lin{0:06d}.png"
.format(fno), img_rsz)
#
print("frame {} is complete.".format(fno))
def main():
args = parse_args()
video_path = args.video_path
weight_path = args.weight_path
if pipeline:
input_queue = JoinableQueue()
pre_process = Process(target=pre_processor,
args=((input_queue, video_path), ))
pre_process.start()
output_queue = SimpleQueue()
post_process = Process(target=post_processor,
args=((output_queue, args.visualize), ))
post_process.start()
else:
cap = cv2.VideoCapture(video_path)
save_dict = torch.load(weight_path, map_location='cpu')
net.load_state_dict(save_dict['net'])
net.eval()
net.cuda()
while True:
if pipeline:
loop_start = time.time()
x = input_queue.get()
input_queue.task_done()
gpu_start = time.time()
seg_pred, exist_pred = network(net, x)
gpu_end = time.time()
output_queue.put((x, seg_pred, exist_pred))
loop_end = time.time()
else:
if not cap.isOpened():
break
ret, frame = cap.read()
if ret:
loop_start = time.time()
frame = transform_img({'img': frame})['img']
img = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
x = transform_to_net({'img': img})['img']
x.unsqueeze_(0)
gpu_start = time.time()
seg_pred, exist_pred = network(net, x)
gpu_end = time.time()
seg_pred = seg_pred.numpy()[0]
exist_pred = exist_pred.numpy()
exist = [1 if exist_pred[0, i] > 0.5 else 0 for i in range(4)]
i2i2 = []
#change
for i in getLane.prob2lines_CULane(seg_pred, exist):
i2i2 += i
pass
loop_end = time.time()
if args.visualize:
img = visualize(img, seg_pred, exist_pred)
cv2.imshow('input_video', frame)
cv2.imshow("output_video", img)
if cv2.waitKey(1) & 0xFF == ord('q'):
break
else:
break
print("gpu_runtime:", gpu_end - gpu_start, "FPS:",
int(1 / (gpu_end - gpu_start)))
print("total_runtime:", loop_end - loop_start, "FPS:",
int(1 / (loop_end - loop_start)))
cv2.destroyAllWindows()
def main():
# ------------ Arguments parsing ------------
args = parse_args()
img_path = args.img_path
weight_path = args.weight_path
exp_dir = args.exp_dir
while exp_dir[-1]=='/':
exp_dir = exp_dir[:-1]
exp_name = exp_dir.split('/')[-1]
with open(os.path.join(exp_dir, "model_config.json")) as f:
exp_cfg = json.load(f)
resize_shape = tuple(exp_cfg['dataset']['resize_shape'])
# net = SCNN(input_size=(512, 288), pretrained=False)
# ------------ preparation ------------
if args.model == "scnn":
net = SCNN(resize_shape, pretrained=False)
elif args.model == "enet_sad":
net = ENet_SAD(resize_shape, sad=True)
else:
raise Exception("Model not match. 'model' in 'model_config.json' should be 'scnn' or 'enet_sad'.")
mean=(0.3598, 0.3653, 0.3662) # TuSimple mean, std
std=(0.2573, 0.2663, 0.2756)
# Resize the image for TuSimple format
transform_img = Resize(resize_shape)
transform_to_net = Compose(ToTensor(), Normalize(mean=mean, std=std))
img = cv2.imread(img_path)
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
img = transform_img({'img': img})['img']
x = transform_to_net({'img': img})['img']
x.unsqueeze_(0)
save_dict = torch.load(weight_path, map_location='cpu')
net.load_state_dict(save_dict['net'])
net.eval()
seg_pred, exist_pred = net(x)[:2]
seg_pred = seg_pred.detach().cpu().numpy()
exist_pred = exist_pred.detach().cpu().numpy()
seg_pred = seg_pred[0]
exist = [1 if exist_pred[0, i] > 0.5 else 0 for i in range(4)]
img = cv2.cvtColor(img, cv2.COLOR_RGB2BGR)
lane_img = np.zeros_like(img)
color = np.array([[255, 125, 0], [0, 255, 0], [0, 0, 255], [0, 255, 255]], dtype='uint8')
coord_mask = np.argmax(seg_pred, axis=0)
for i in range(0, 4):
if exist_pred[0, i] > 0.5:
lane_img[coord_mask == (i + 1)] = color[i]
img = cv2.addWeighted(src1=lane_img, alpha=0.8, src2=img, beta=1., gamma=0.)
cv2.imwrite("image/demo_result.jpg", img)
for x in getLane.prob2lines_CULane(seg_pred, exist):
print(x)
if args.visualize:
print([1 if exist_pred[0, i] > 0.5 else 0 for i in range(4)])
cv2.imshow("", img)
cv2.waitKey(0)
cv2.destroyAllWindows()
img = transform({'img': img})['img']
img = img.unsqueeze(0).to(device)
with torch.no_grad():
seg_pred, exist_pred = net(img)[:2]
seg_pred = F.softmax(seg_pred, dim=1)
seg_pred = seg_pred.cpu().numpy()
exist_pred = exist_pred.cpu().numpy()
b=0
seg = seg_pred[b]
exist = [1 if exist_pred[b, i] > 0.5 else 0 for i in range(4)]
if dataset_name == 'Tusimple':
lane_coords = getLane.prob2lines_tusimple(seg, exist, resize_shape=original_shape[::-1], y_px_gap=10, pts=56)
elif dataset_name == 'CULane':
lane_coords = getLane.prob2lines_CULane(seg, exist, resize_shape=original_shape[::-1], y_px_gap=20, pts=18)
for i in range(len(lane_coords)):
lane_coords[i] = sorted(lane_coords[i], key=lambda pair: pair[1])
for lane in lane_coords:
foo = np.int32([lane])[:,np.int32([lane])[0,:,0] >= 0]
cv2.polylines(img_vis, foo, isClosed=False, color=(0,0,255), thickness=2)
#show segmentation map
seg = seg.swapaxes(0,2).swapaxes(0,1)
seg = cv2.resize(seg, dsize=original_shape, interpolation=cv2.INTER_CUBIC)
lane_img = np.zeros_like(img_vis)
color = np.array([[255, 125, 0], [0, 255, 0], [0, 0, 255], [0, 255, 255]], dtype='uint8')
coord_mask = np.argmax(seg, axis=-1)
for i in range(0, 4):
def main():
args = parse_args()
img_path = args.img_path
weight_path = args.weight_path
img = cv2.imread(img_path)
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
x = transform({'img': img})['img']
x.unsqueeze_(0)
save_dict = torch.load(weight_path, map_location='cpu')
net.load_state_dict(save_dict['net'])
net.eval()
if torch.cuda.is_available():
x = x.cuda()
# start_time = time.time()
# for i in tqdm(range(100)):
# seg_pred, exist_pred = net(x)[:2]
# print("Spend time {} seconds for 100 times".format(time.time() - start_time))
# CPU spend time 64.43067598342896 seconds for 100 times, --> 640ms
# GPU spend time 7.734357118606567 seconds for 100 times, --> 70ms
seg_pred, exist_pred = net(x)[:2]
# pdb.set_trace()
# (pdb) pp x.shape
# torch.Size([1, 3, 288, 512])
# (pdb) pp seg_pred.shape
# torch.Size([1, 5, 288, 512])
# (Pdb) pp seg_pred.min(), seg_pred.max()
# (tensor(-13.8127, device='cuda:0', grad_fn=<MinBackward1>),
# tensor(26.3522, device='cuda:0', grad_fn=<MaxBackward1>))
# (Pdb) pp exist_pred.shape, exist_pred
# (torch.Size([1, 4]),
# tensor([[0.9989, 0.9999, 0.9993, 0.9983]], device='cuda:0',
# grad_fn=<SigmoidBackward>))
seg_pred = seg_pred.detach().cpu().numpy()
exist_pred = exist_pred.detach().cpu().numpy()
seg_pred = seg_pred[0]
exist = [1 if exist_pred[0, i] > 0.5 else 0 for i in range(4)]
img = cv2.cvtColor(img, cv2.COLOR_RGB2BGR)
img = cv2.resize(img, (image_resize_width, 288))
lane_img = np.zeros_like(img)
color = np.array([[255, 125, 0], [0, 255, 0], [0, 0, 255], [0, 255, 255]], dtype='uint8')
coord_mask = np.argmax(seg_pred, axis=0)
for i in range(0, 4):
if exist_pred[0, i] > 0.5:
lane_img[coord_mask == (i + 1)] = color[i]
cv2.imshow("lane", lane_img)
# pdb.set_trace()
# (Pdb) seg_pred.shape, coord_mask.shape
# ((5, 288, 512), (288, 512))
img = cv2.addWeighted(src1=lane_img, alpha=0.8, src2=img, beta=1., gamma=0.)
for x in getLane.prob2lines_CULane(seg_pred, exist):
print(x)
if args.visualize:
print([1 if exist_pred[0, i] > 0.5 else 0 for i in range(4)])
cv2.imshow("", img)
cv2.waitKey(0)
cv2.destroyAllWindows()
else:
cv2.imwrite("demo/demo_result.jpg", img)
def main():
args = parse_args()
video_path = args.video_path
weight_path = args.weight_path
# Loading configs
exp_dir = args.exp_dir
while exp_dir[-1] == '/':
exp_dir = exp_dir[:-1]
exp_name = exp_dir.split('/')[-1]
with open(os.path.join(exp_dir, "model_config.json")) as f:
exp_cfg = json.load(f)
resize_shape = tuple(exp_cfg['dataset']['resize_shape'])
# Build the scnn/enet-sad model according to the argument
if args.model == "scnn":
net = SCNN(resize_shape, pretrained=True)
elif args.model == "enet_sad":
net = ENet_SAD(resize_shape, sad=True)
else:
raise Exception(
"Model not match. '--model' in argument should be 'scnn' or 'enet_sad'."
)
mean = (0.485, 0.456, 0.406)
std = (0.229, 0.224, 0.225)
# Resize the image for TuSimple format
transform_img = Resize(resize_shape)
transform_to_net = Compose(ToTensor(), Normalize(mean=mean, std=std))
if pipeline:
input_queue = JoinableQueue()
pre_process = Process(target=pre_processor,
args=((input_queue, video_path), ))
pre_process.start()
output_queue = SimpleQueue()
post_process = Process(target=post_processor,
args=((output_queue, args.visualize), ))
post_process.start()
else:
cap = cv2.VideoCapture(video_path)
save_dict = torch.load(weight_path, map_location='cpu')
net.load_state_dict(save_dict['net'])
net.eval()
net.cuda()
result = cv2.VideoWriter('video/demo.avi', cv2.VideoWriter_fourcc(*'MJPG'),
30, resize_shape)
while True:
if pipeline:
loop_start = time.time()
x = input_queue.get()
input_queue.task_done()
gpu_start = time.time()
seg_pred, exist_pred = network(net, x)
gpu_end = time.time()
output_queue.put((x, seg_pred, exist_pred))
loop_end = time.time()
else:
if not cap.isOpened():
break
ret, frame = cap.read()
if ret:
loop_start = time.time()
#frame = cv2.rotate(frame, cv2.ROTATE_90_COUNTERCLOCKWISE)
frame = transform_img({'img': frame})['img']
img = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
x = transform_to_net({'img': img})['img']
x.unsqueeze_(0)
gpu_start = time.time()
seg_pred, exist_pred = network(net, x)
gpu_end = time.time()
seg_pred = seg_pred.numpy()[0]
exist_pred = exist_pred.numpy()
exist = [1 if exist_pred[0, i] > 0.5 else 0 for i in range(4)]
print(exist)
for i in getLane.prob2lines_CULane(seg_pred, exist):
print(i)
loop_end = time.time()
if args.visualize:
img = visualize(img, seg_pred, exist_pred)
#cv2.imshow('input_video', frame)
#cv2.imshow("output_video", img)
result.write(img)
if cv2.waitKey(1) & 0xFF == ord('q'):
break
else:
break
print("gpu_runtime:", gpu_end - gpu_start, "FPS:",
int(1 / (gpu_end - gpu_start)))
print("total_runtime:", loop_end - loop_start, "FPS:",
int(1 / (loop_end - loop_start)))
result.release()
cv2.destroyAllWindows()
