Exploding REL value about sparse-to-dense.pytorch

[timethy]

While this is minor, it sometimes results in a
RuntimeError: Overflow when unpacking long

while writing into the csv which aborts training.

[fangchangma]

I understand the problem you described. However, I would suggest first making sure your data generation process is correct. Specifically, unlike a typical color image (which is dense), resizing sparse depth images can lead to spurious pixel values due to the bi-linear interpolation. 1e-10 is unlikely to be the real ground truth.

If possible, try to avoid resizing sparse depth image at all. If resizing is indeed necessary, use nearest neighbor for interpolation, rather than bi-linear or cubic interpolation.

[timethy]

Hi, yes, excactly. I observed that this also happens with the NYUDepthv2 dataset, so maybe we should add a filter there to get rid of these small depth values.

[fangchangma]

The culprit of the exploding REL values is a bug in the rotation (I was using bilinear interpolation rather than nearest neighbor). This has been fixed and the REL value remains stable during training.