image, radii = rasterizer() but get all-zero image, radii

[nanasylum]

Hi, congratulations for ur great work!!!

I tried to apply mvsplat on other datasets, but encountered a problem. I debug and find that in src.model.decoder.cuda_splatting.py line118
`

        image, radii = rasterizer(

        means3D=gaussian_means[i],

        means2D=mean_gradients,

        shs=shs[i] if use_sh else None,

        colors_precomp=None if use_sh else shs[i, :, 0, :],
        
        opacities=gaussian_opacities[i, ..., None],

        cov3D_precomp=gaussian_covariances[i, :, row, col],
    )

`
I got all-zero image and radii.
Here is my other parameter:

( I guess )maybe my parameter is sooo small so that it results overflow???
I dont know why, could you plz help me?
btw, i wanna know how rasterizer() works, but i dont know how to debug in .cu files.

[donydchen]

Hi @nanasylum, thanks for your appreciation.

May I confirm whether you can get the correct rendered outputs when testing with the released pre-trained weight? If you still get black images (all zero), try to manually re-install the 3DGS renderer by git clone --recursive the diff-gaussian-rasterization-modified project, see #60.

If it works correctly, try debugging through camera poses, near and far depth scales, image resolutions, etc. For more detailed instructions, see #23 (comment).

The released RE10K pre-trained weight should produce reasonable outputs on several datasets. If none of the above approaches helps, let me know your dataset and how you obtain the camera poses, e.g., using COLMAP. Have fun tuning.