ShijieZhou-UCLA / feature-3dgs

[CVPR 2024 Highlight] Feature 3DGS: Supercharging 3D Gaussian Splatting to Enable Distilled Feature Fields
Other
326 stars 21 forks source link

Issue about render feature #30

Closed Doctor-James closed 3 months ago

Doctor-James commented 3 months ago

Hello, thank you for your excellent work. I have encountered some issues while porting your diff-gaussian-rasterization to my project:

RuntimeError: Function _RasterizeGaussiansBackward returned an invalid gradient at index 4 - got [4224, 256, 256] but expected shape compatible with [4224, 256]

rendered_image, rendered_feature, radii, depth = rasterizer( means3d, # 4224, 3 means2d, # 4224, 3 opacity, # 4224, 1 colors_precomp=colors, # 4224, 3 semantic_feature = feature, # 4224, 256 scales=scales, # 4224, 3 rotations=rotations) # 4224, 4

It seems that the forward propagation is working fine, but there is an issue during the gradient backpropagation. Since I am not familiar with CUDA code, I would like to ask if you have any suggestions or debugging methods.

yzslab commented 3 months ago

Before getting a answer from the author. I think replace with gsplat can be a temporary solution. Here is a patch can be apply to this repo. directly: https://github.com/ShijieZhou-UCLA/feature-3dgs/issues/14#issuecomment-2185639700

Doctor-James commented 3 months ago

Thanks,it works