def testone():
# for param in model.parameters():
# print(param.data)
line = '000016'
print("Img = ", line)
x, y, h, w, _ = load_data(line, testing=True)
y_pred = model(x)
loss_fn = torch.nn.MSELoss(reduction='mean')
loss = loss_fn(w * y_pred, w * y)
print("loss = ", loss)
# print(x)
# print(y_pred)
y_pred = y_pred.cpu()
pred_img = np.reshape(y_pred.detach().numpy(), (h, w))
disp = pred_img
pred_img = (256 * pred_img).astype('uint16')
# print(pred_img)
# print(np.max(pred_img))
im = Image.fromarray(pred_img)
directory = './disp_1d_w/'
if not path.exists(directory):
os.makedirs(directory)
im.save(directory + line + '.png')
calib_file = CALIB_PATH + line + '.txt'
calib = kitti_util.Calibration(calib_file)
max_high = 1
lidar = generate_lidar.project_disp_to_points(calib, disp, max_high)
lidar = np.concatenate([lidar, np.ones((lidar.shape[0], 1))], 1)
lidar = lidar.astype(np.float32)
lidar.tofile('{}/{}.bin'.format('./pl/', line))
def test_one(index):
disp = imageio.imread(DISPARITY_PATH + index + '.png') / 256
# disp = filter_one(disp)
disp = test_mask(index, disp)
#disp = semantic_mask(index, disp)
# disp = filter_one(disp)
print("Saved Disparity")
imageio.imwrite('./filtered_disp/' + index + '.png',
(disp * 256).astype('uint16'))
calib_file = CALIB_PATH + index + '.txt'
calib = kitti_util.Calibration(calib_file)
max_high = 1
lidar = generate_lidar.project_disp_to_points(calib, disp, max_high)
lidar = np.concatenate([lidar, np.ones((lidar.shape[0], 1))], 1)
lidar = lidar.astype(np.float32)
lidar.tofile('{}/{}.bin'.format('./pl/', index))
print("Saved point cloud")
def semantic_filter_all():
file1 = open('../KITTI/object/trainval.txt', 'r')
lines = file1.read().splitlines()
file1.close()
if not os.path.exists(SAVE_PL_PATH):
os.makedirs(SAVE_PL_PATH)
for line in lines:
index = line
disp = imageio.imread(DISPARITY_PATH + index + '.png') / 256
# disp = obj_mask(index, disp)
disp = dedge_obj_mask(index, disp)
calib_file = CALIB_PATH + index + '.txt'
calib = kitti_util.Calibration(calib_file)
max_high = 1
lidar = generate_lidar.project_disp_to_points(calib, disp, max_high)
lidar = np.concatenate([lidar, np.ones((lidar.shape[0], 1))], 1)
lidar = lidar.astype(np.float32)
lidar.tofile(SAVE_PL_PATH + index + '.bin')
print("Saved point cloud # " + index)
def depth_to_disp():
DEPTH_PATH = '/home/maxc303/aer1515/ip_basic/demos/outputs/kitti_depth_ipbasic/'
file1 = open('../KITTI/object/train.txt', 'r')
lines = file1.read().splitlines()
file1.close()
if not os.path.exists(SAVE_PL_PATH):
os.makedirs(SAVE_PL_PATH)
for line in lines:
index = line
depth = imageio.imread(DEPTH_PATH + index + '.png') / 256
# disp = obj_mask(index, disp)
calib_file = CALIB_PATH + index + '.txt'
calib = kitti_util.Calibration(calib_file)
max_high = 1
baseline = 0.54
depth[depth == 0] = -1
disp = (calib.f_u * baseline) / depth
disp[disp < 0] = 0
disp[disp > 192] = 0
imageio.imwrite(SAVE_DISP_PATH + index + '.png',
(disp * 256).astype('uint16'))
print("Saved depth # " + index)
def run_from_file(self):
if self.args_disp.KITTI == '2015':
from psmnet.dataloader import KITTI_submission_loader as DA
else:
from psmnet.dataloader import KITTI_submission_loader2012 as DA
test_left_img, test_right_img = DA.dataloader(self.args_disp.datapath)
if not os.path.isdir(self.args_disp.save_path):
os.makedirs(self.args_disp.save_path)
# vis = o3d.visualization.Visualizer()
# vis.create_window()
# pcd = o3d.geometry.PointCloud()
# pcd.points = o3d.utility.Vector3dVector(np.ones([100,3]))
# vis.add_geometry(pcd)
for inx in range(len(test_left_img)):
################ part 1 #######################
imgL_o = (skimage.io.imread(test_left_img[inx]).astype('float32'))
imgR_o = (skimage.io.imread(test_right_img[inx]).astype('float32'))
img = self.disp_pred_net.run(imgL_o, imgR_o)
#####visualize
cv2.imshow('disparity image', img.astype(np.uint8))
cv2.waitKey(1)
##################
# # file output
# print(test_left_img[inx].split('/')[-1])
# if self.args.save_figure:
# skimage.io.imsave(self.args.save_path+'/'+test_left_img[inx].split('/')[-1],(img*256).astype('uint16'))
# else:
# np.save(self.args.save_path+'/'+test_left_img[inx].split('/')[-1][:-4], img)
################# part 2 ###################
predix = test_left_img[inx].split('/')[-1][:-4]
calib_file = '{}/{}.txt'.format(self.args_gen_lidar.calib_dir,
predix)
calib = kitti_util.Calibration(calib_file)
img = (img * 256).astype(np.uint16) / 256.
lidar, valid_idx = self.pcl_generator.run(calib, img)
# pad 1 in the indensity dimension
# lidar = np.concatenate([lidar, np.ones((lidar.shape[0], 1))], 1)
lidar = lidar.astype(np.float32) # from float64 to float32
# lidar.tofile('{}/{}.bin'.format(self.args_gen_lidar.save_dir, predix))
print('Finish Depth {}'.format(predix))
##### visualize
# pcd.points = o3d.utility.Vector3dVector(lidar)
imgL_o_flat = imgL_o.reshape([-1, 3])
imgL_o_valid = imgL_o_flat[valid_idx] / 255.
# print(imgL_o_valid)
# print(imgL_o_valid.shape)
# print(lidar.shape)
pcd = o3d.geometry.PointCloud()
pcd.points = o3d.utility.Vector3dVector(lidar)
pcd.colors = o3d.utility.Vector3dVector(imgL_o_valid)
# pcd.paint_uniform_color([1, 0.706, 0])
o3d.visualization.draw_geometries([pcd])
# vis.update_geometry()
# vis.poll_events()
# vis.update_renderer()
##################### part 3 ############################
vis.destroy_window()
lidar.tofile(path)
return path
if __name__ == "__main__":
if len(sys.argv) < 2:
print('Usage ./exe 000025')
sys.exit(0)
img_file = sys.argv[1] + '.png'
calib_file = sys.argv[1] + '.txt'
depth_file = sys.argv[1] + '.png.npy'
seg_file = sys.argv[1] + 's.png'
gt_lidar_file = sys.argv[1] + 'gt.bin'
image = cv2.imread(img_file)
calib = kitti_util.Calibration(calib_file)
depth_map = np.load(depth_file)
seg_map = cv2.imread(seg_file, 0)
seg_mask = (seg_map==13).astype(np.uint8)
kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (10, 10)) # 椭圆
seg_mask = cv2.erode(seg_mask, kernel) > 0
gt_lidar = np.fromfile(gt_lidar_file, dtype=np.float32).reshape(-1, 4)
proj_lidar = project_disp_to_depth(calib, depth_map, seg_mask)
# pad 1 in the indensity dimension
proj_lidar = np.concatenate([proj_lidar, np.ones((proj_lidar.shape[0], 1))], 1)
out_lidar = np.concatenate([gt_lidar, proj_lidar], axis=0)
f = to_bin_file(out_lidar, sys.argv[1]+'p')
print('Finish Depth {} {}'.format(sys.argv[1], out_lidar.shape[0]))