nerfstudio-project
Taming gs for faster training and rendering
Taming 3DGS released an optimized 3dgs implementation which seems to be faster than the current gsplat. The released code mostly related to section 4 of their paper. Maybe similar strategy can be considered in gsplat.
https://github.com/nullptr81/3dgs-accel
their backward implementation is slower than simple warp-reduce backward in my codebase (not gsplat, backward algo is similar), don't need to waste time on that. three code segments that have speed up in this paper:
- tighter culling
- fused ssim
- split sh coeffs to two tensor to avoid expensive concat.
I think Taming-gs's biggest contribution is its densification strategy, and I hope gsplat can add support for this strategy in the future.
@rahul-goel Are you planning on adding Taming-gs's densification strategy to gsplat?
Hi @xiemeilong. I don't think there are any plans to add this to gsplat. The code will be released by November end here: https://github.com/humansensinglab/taming-3dgs.
their backward implementation is slower than simple warp-reduce backward in my codebase (not gsplat, backward algo is similar), don't need to waste time on that. three code segments that have speed up in this paper:
- tighter culling
- fused ssim
- split sh coeffs to two tensor to avoid expensive concat.
Do you have any plans to release your more efficient backward version?
Graphdeco-inria has adopted 3dgs_accel or taming-gs rasterizer implementation in this branch .
Can these idea be integrated into gsplat for further accelerating. https://github.com/humansensinglab/taming-3dgs/blob/main/submodules/diff-gaussian-rasterization/cuda_rasterizer/forward.cu#L349 https://github.com/humansensinglab/taming-3dgs/blob/main/submodules/diff-gaussian-rasterization/cuda_rasterizer/backward.cu#L404
