search

autonomousvision / gaussian-opacity-fields

Gaussian Opacity Fields: Efficient and Compact Surface Reconstruction in Unbounded Scenes
https://niujinshuchong.github.io/gaussian-opacity-fields/
Other
559 stars 26 forks source link

RuntimeError: CUDA error: an illegal memory access was encountered #31

Closed mudimingquedeyinmoujia closed 1 month ago

mudimingquedeyinmoujia commented 1 month ago

Hi, I tried to modify the codes of Gaussians Representation while keeping your rasterizer parts (for the using of normal map and depth map). It works well at first, but I got: RuntimeError: CUDA error: an illegal memory access was encountered at the 114 line of your gaussian-opacity-fields/gaussian_renderer/init.py, that is : "visibility_filter" : radii > 0

This problem occurs at the iteration of about 2k-5k, and it seems like a metaphysical problem. I can not fingure out why. So I doubt are there any rubust problem in the CUDA codes?

huahangc commented 1 month ago

I also get this error which I use the GOF rasterization in my own dataset.

niujinshuchong commented 1 month ago

Could you check if you are already using this one https://github.com/graphdeco-inria/gaussian-splatting/issues/41#issuecomment-1752279620 ? Or can you be more specific for what you have changed and what kind of data you are using?

mudimingquedeyinmoujia commented 1 month ago

@niujinshuchong Hi, thanks for your reply. I solved the problem. This is because there are NAN input of gaussian rasterizer while the rasiterizer can not handle the situation. @huahangc I guess maybe there are NAN of your data or GS, so carefully check the NAN value before the rasterizer.

huahangc commented 1 month ago

@mudimingquedeyinmoujia Thanks. There does has nan value input of the gaussian rasterizer. But all loss turn to nan when I check the nan value of the input and turn the nan value to 0. How to solve this problem?