Reproduce “LiteGS: A High-performance Framework to Train 3DGS in Subminutes via System and Algorithm Codesign” 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
Kaimin Liao, Hua Wang, Zhi Chen, Luchao Wang, Yaohua Tang
LiteGS trains 3D Gaussian Splatting scenes from photographs for new views. It reorganizes GPU work and scene data, and adds primitives where pixel-level opacity-gradient variation suggests more detail. On Mip-NeRF 360, it trains in 301 seconds with 1.10 million Gaussians, versus 3,142 seconds and 3.21 million for 3DGS-MCMC. PSNR is similar, but SSIM and LPIPS are worse.
Key insights
- A holistic "system and algorithm codesign" approach, integrating optimizations across GPU computation, data management, and algorithmic layers, yields substantial synergistic performance gains.
- Drastically reducing gradient reduction overheads through warp-based rasterization, scanline algorithms, and mixed-precision computing is fundamental for high GPU utilization.
- Dynamic spatial sorting via Morton codes combined with cluster-level culling and compaction effectively addresses scalability issues from disordered primitives and poor cache locality.
- Replacing position-gradient-based densification with an opacity-gradient-variance metric and stable opacity decay leads to more robust, precise, and high-quality scene parameter growth.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.