Reproduce “Speeding Up the Learning of 3D Gaussians with Much Shorter Gaussian Lists” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jiaqi Liu, Zhizhong Han
The authors accelerate 3D Gaussian splatting training by shortening per-pixel Gaussian lists, rather than reducing the scene’s total Gaussian count. They periodically shrink Gaussian scales and sharpen alpha-blending contributions through entropy regularization, alongside a resolution scheduler and LiteGS. On Mip-NeRF 360, training takes 99.58 seconds versus 919.51 for 3DGS, with modestly lower PSNR; the full speedup is not attributable to shorter lists alone.
Key insights
- Directly reducing the length of per-pixel Gaussian lists is a highly effective approach to accelerate 3DGS training, complementing other optimization strategies.
- Periodic scale reduction offers a more direct and stable method for controlling Gaussian size and influence compared to continuous volume regularization.
- Minimizing the entropy of blending weights along a ray encourages Gaussians to localize their contributions, leading to more focused representation and shorter per-pixel lists.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.