Issue related to point cloud generation from the ground-truth image in the dataloader during training and in vis_data.py

[SMohammadi89]

Hi, thanks for releasing the training code!

I'm trying to finetune the model on the TartanAir dataset. However, when I visualize the reconstructed point cloud using vis_data.py, the output appears as a single line instead of a proper 3D structure. This issue also occurs during training.

Interestingly, when I use a manual function to reconstruct the point cloud, it appears correctly. Do you have any insights on what might be causing this issue?

[XuanchenLi]

I have the same problem. Do you have any ideas?

[XuanchenLi]

I figured it out, the intrinsics in moge/utils3d need to be normalized. The cx and cy should be 0.5. and the fx, fy should be divided by the image size: fx = fx / w; fy = fy / h

[SMohammadi89]

@XuanchenLi Thank you so much for the help