def test_single_image(img_path, model, training_hw, save_dir='./'):
img = cv2.imread(img_path)
h, w = img.shape[0:2]
img_resized = cv2.resize(img, (training_hw[1], training_hw[0]))
img_t = torch.from_numpy(np.transpose(
img_resized, [2, 0, 1])).float().cuda().unsqueeze(0) / 255.0
disp = model.infer_depth(img_t)
disp = np.transpose(disp[0].cpu().detach().numpy(), [1, 2, 0])
disp_resized = cv2.resize(disp, (w, h))
depth = 1.0 / (1e-6 + disp_resized)
visualizer = Visualizer_debug(dump_dir=save_dir)
visualizer.save_disp_color_img(disp_resized, name='demo')
print('Depth prediction saved in ' + save_dir)
def test_kitti_2015(cfg,
model,
gt_flows,
noc_masks,
gt_masks,
depth_save_dir=None):
dataset = KITTI_2015(cfg.gt_2015_dir)
visualizer = Visualizer_debug(depth_save_dir)
pred_flow_list = []
pred_disp_list = []
img_list = []
for idx, inputs in enumerate(tqdm(dataset)):
img, K, K_inv = inputs
img = img[None, :, :, :]
K = K[None, :, :]
K_inv = K_inv[None, :, :]
img_h = int(img.shape[2] / 2)
img1, img2 = img[:, :, :img_h, :], img[:, :, img_h:, :]
img_list.append(img1)
img1, img2, K, K_inv = img1.cuda(), img2.cuda(), K.cuda(), K_inv.cuda()
if cfg.mode == 'flow' or cfg.mode == 'flowposenet':
flow = model.inference_flow(img1, img2)
else:
flow, disp1, disp2, Rt, _, _ = model.inference(
img1, img2, K, K_inv)
disp = disp1[0].detach().cpu().numpy()
disp = disp.transpose(1, 2, 0)
pred_disp_list.append(disp)
flow = flow[0].detach().cpu().numpy()
flow = flow.transpose(1, 2, 0)
pred_flow_list.append(flow)
# pdb.set_trace()
eval_flow_res = eval_flow_avg(gt_flows,
noc_masks,
pred_flow_list,
cfg,
moving_masks=gt_masks,
write_img=False)
print('CONFIG: {0}, mode: {1}'.format(cfg.config_file, cfg.mode))
print('[EVAL] [KITTI 2015]')
print(eval_flow_res)
# depth evaluation
return eval_flow_res