Renderer may have truncation error

[clmforgit]

hi, dear author,

Thanks for your excellent work! I test the renderer, using the code like this:

 for i in range(5):
      cameras = PerspectiveCameras(
          R=R_pytorch3d,
          T=T_pytorch3d + torch.Tensor([0,0,i * 0.05]).type_as(T_pytorch3d),
          focal_length=focal_pytorch3d,
          principal_point=p0_pytorch3d,
          image_size=image_shape,
          device=device
      )

so the rendered result should move slightly. But actually, some images are exactly the same, which is not as expected. My pytorch3d verson is 0.6.1. Can you reproduce this?

[DRosemei]

@clmforgit Can you provide some gt images and render images?

[clmforgit]

@clmforgit Can you provide some gt images and render images?

the 5 blend images are generated using eval.py from a grid model, by modifying code of cameras to get 5 rendered seg for 1 gt image. It can be seen that the 2nd 3rd 4th images are exactly the same.

[DRosemei]

@clmforgit The first and last image are not the same in details. Moving 5cm for each image makes little changes.

[DRosemei]

@clmforgit The key reason is that your bev resolution is set to 10cm, if you want to see more details, decrease the bev resolution.

[clmforgit]

Yes, that makes sense. I doubt if it is a bug or something in pytorch3d.