encountered an AssertionError while using UniDepthV2 to predict depth

[red-liu]

I really appreciate your great masterpiece. but I used UniDepthV2 to predict, encountered an AssertionError exception as below:

 File "/home/user/app/app.py", line 23, in <module>
 predictions = model.infer(rgb)
File "/home/user/.pyenv/versions/3.10.14/lib/python3.10/site-packages/torch/utils/_contextlib.py", line 115, in decorate_context
return func(*args, **kwargs)
File "/home/user/app/unidepth/models/unidepthv2/unidepthv2.py", line 229, in infer
features, tokens = self.pixel_encoder(rgbs)
File "/home/user/.pyenv/versions/3.10.14/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1511, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
File "/home/user/.pyenv/versions/3.10.14/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1520, in _call_impl
return forward_call(*args, **kwargs)
File "/home/user/app/unidepth/models/backbones/dinov2.py", line 324, in forward
x = self.prepare_tokens_with_masks(x, masks)
File "/home/user/app/unidepth/models/backbones/dinov2.py", line 312, in prepare_tokens_with_masks
x = x + self.interpolate_pos_encoding(x, w, h)
File "/home/user/app/unidepth/models/backbones/dinov2.py", line 297, in interpolate_pos_encoding
and int(h0) == patch_pos_embed.shape[-1]

my result is: int(w0)= 57,patch_pos_embed.shape[-2]= 57 and int(h0)= 43,patch_pos_embed.shape[-1]= 42

[red-liu]

another a question：if I tranform a picture to less resolution or more resolution，how will the result change？

[lpiccinelli-eth]

Thanks for using our work!

Which is your input shape? (or the config you are passing to the mode, like pixels_bounds, etc..)

To answer your question: the results may change a bit, but we expects them to be quite consistent, something that is not typical for previous works, especially in case of metric estimation.

[red-liu]

Thank you very much for your reply.
My input is a picture, its shape is (4032, 3024) and it is from iphone 13, so height is bigger than width.
pixels_bounds had no specific setting because I don't understand the purpose.

[BaderTim]

hi there, I encounter a similar error with KITTI-shaped images.

[lpiccinelli-eth]

Thank you for the info, it looks like when out of bounds of the ratio, it fails, I will check it and get back to you (hopefully) the corrected version.

[lpiccinelli-eth]

The error comes from DINO original code and was solved in this PR, we committed the changes and now it should be solved.

Let me know if something is still off.

[red-liu]

It seems the issue has been resolved. Thank you very much for your help. By the way, I've encountered a new problem: there's a significant difference between the intrinsics predictions from version 1 and version 2. Do you have any idea what might be causing this discrepancy?
To the same picture, the intrinsics predictions as below:

v2:
    [[[3.8631e+03, 0.0000e+00, 1.5201e+03],
     [0.0000e+00, 4.0359e+03, 2.0109e+03],
     [0.0000e+00, 0.0000e+00, 1.0000e+00]]]

v1:
 [[[1.6742e+03, 0.0000e+00, 1.5174e+03],
     [0.0000e+00, 2.7255e+03, 2.0222e+03],
     [0.0000e+00, 0.0000e+00, 1.0000e+00]]]