Add interface for depth in both forward rendering and backward propagation

[ingra14m]

Test in self-defined dataset with GT-depth

Without depth loss:

With depth loss

[XuyangBai]

Should there be a second part of gradient affected by the depth loss, the dL_dtz in L264 is also affected by the dL_depths?

[cv-lab-x]

Test in self-defined dataset with GT-depth Without depth loss:

With depth loss

Hi，thanks for your work, have you transformed the render depths to point3ds? I test your latest branch on datasets bicycle mip360-datasets, it seems the render depths have some erros, and multi views's render depths are not consistent. Looking forward to your reply, thanks! @ingra14m

[shippoT]

Hi，how do you get the depth map shown in ”Test in self-defined dataset with GT-depth“. I rendered the same NeRF Synthetic dataset using blender，but got quite different result.

[ingra14m]

If I remember correctly, first, I set the parts where alpha=0 to be black.

Second, since the depth directly from Blender shows deeper depths in blacker colors, I used 1.0 - normalized depth. In this way, the depth gt can be directly used to supervise the depth directly output by gs.

[arcman7]

Hi，thanks for your work, have you transformed the render depths to point3ds? I test your latest branch on datasets bicycle mip360-datasets, it seems the render depths have some erros, and multi views's render depths are not consistent. Looking forward to your reply, thanks! @ingra14m

What's the status of this branch? Does including the depth information improve the splat fitting or rendering?

[ingra14m]

Hi @arcman7, from my perspective, I think depth can not improve the rendering quality of 3D Gaussian splatting. The geometry enhancement cannot lead to a better rendering quality.

[arcman7]

Hi @arcman7, from my perspective, I think depth can not improve the rendering quality of 3D Gaussian splatting. The geometry enhancement cannot lead to a better rendering quality.

Ah okay, thanks for the heads up - it was looking really hopeful when I was going through all of the related discussion threads and experiments people had setup

[arcman7]

Hey btw, I wanted to ask you a quick question, and I didn't think creating another issue on this busy repo was the way to go; is there a way to render a pretrained gaussian model without setting up training gradient tensors in cuda? I'm not trying to improve or modify this pre-existing gaussian splat model, I just want to rasterize it from it's various camera view points but using the existing py-torch rasterization methods in this repo

[zcc00210]

Test in self-defined dataset with GT-depth Without depth loss:
With depth loss

Hi，thanks for your work, have you transformed the render depths to point3ds? I test your latest branch on datasets bicycle mip360-datasets, it seems the render depths have some erros, and multi views's render depths are not consistent. Looking forward to your reply, thanks! @ingra14m

Hi,
May I ask if the depth problem of this branch has been solved? I also encountered the problem that the multi-frame view cannot be aligned after being projected to 3d.@cv-lab-x

[zcc00210]

Test in self-defined dataset with GT-depth Without depth loss:

With depth loss

Hi,
May I ask why after I use your depth visualization cuda code under this branch, the rendered depth map output is all 255,255,255?
Your prompt reply will be highly appreciated!@ingra14m