test_disp.py min/max depth value issue

[seb-le]

Hi! Thank you for your nice implementation.

I have a question about clipping the depth value in test_disp.py. https://github.com/ClementPinard/SfmLearner-Pytorch/blob/master/test_disp.py

Currently, before applying the scale factor to predicted depth, depth is clipped by min/max depth value as shown below.

pred_depth_zoomed = zoom(pred_depth,
                         (gt_depth.shape[0]/pred_depth.shape[0],
                          gt_depth.shape[1]/pred_depth.shape[1])
                         ).clip(args.min_depth, args.max_depth)

( ... )

if seq_length > 1:
        ( ... )
        scale_factor = np.mean(sample['displacements'] / displacement_magnitudes)
        errors[0,:,j] = compute_errors(gt_depth, pred_depth_zoomed*scale_factor)

( ... )

scale_factor = np.median(gt_depth)/np.median(pred_depth_zoomed)
errors[1,:,j] = compute_errors(gt_depth, pred_depth_zoomed*scale_factor)

But isn't it correct to clipping after applying the scale factor as shown below?

pred_depth_zoomed = zoom(pred_depth,
                         (gt_depth.shape[0]/pred_depth.shape[0],
                          gt_depth.shape[1]/pred_depth.shape[1])
                         )

( ... )

if seq_length > 1:
        ( ... )
        scale_factor = np.mean(sample['displacements'] / displacement_magnitudes)
        errors[0,:,j] = compute_errors(gt_depth, (pred_depth_zoomed*scale_factor).clip(args.min_depth, args.max_depth))

( ... )

scale_factor = np.median(gt_depth)/np.median(pred_depth_zoomed)
errors[1,:,j] = compute_errors(gt_depth, (pred_depth_zoomed*scale_factor).clip(args.min_depth, args.max_depth))

The difference by this issue is below when pose network is not used.		abs_diff	abs_rel	sq_rel	rms	log_rms	abs_log	a1	a2	a3
current	2.3598	0.1233	0.8640	4.8908	0.1987	0.1258	0.8569	0.9561	0.9815
if fixed	2.3391	0.1229	0.8205	4.7596	0.1981	0.1256	0.8572	0.9564	0.9817

Especially, the sq_rel and rms metric is sensitive to this issue.

Thank you.

[ClementPinard]

Hi, indeed you are right !

Thanks for this catch ! If you submit a PR I'd happily merge it.

[seb-le]

Thank you for your quick reply!

Ok I see. I will submit the PR to your repo.

I also saw this issue in loss_functions.py as shown below. https://github.com/ClementPinard/SfmLearner-Pytorch/blob/master/loss_functions.py

@torch.no_grad()
def compute_depth_errors(gt, pred, crop=True):
        ( ... )
        valid_pred = current_pred[valid].clamp(1e-3, 80)

        valid_pred = valid_pred * torch.median(valid_gt)/torch.median(valid_pred)

Is that also correct like below?

@torch.no_grad()
def compute_depth_errors(gt, pred, crop=True):
        ( ... )
        valid_pred = current_pred[valid]

        valid_pred = (valid_pred * torch.median(valid_gt)/torch.median(valid_pred)).clamp(1e-3, 80)

Thank you!

[ClementPinard]

Yes, the problem is there as well. I will validate this change in your PR :)