Reproduce “Turbo-GS: Accelerating 3D Gaussian Fitting for High-Quality Radiance Fields” 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
Ankit Dhiman, Tao Lu, R Srinath, Emre Arslan, Angela Xing, Yuanbo Xiangli, R Venkatesh Babu, Srinath Sridhar
Turbo-GS introduces a learning-free approach that accelerates 3D Gaussian Splatting (3DGS) model fitting for high-resolution radiance fields. The method reduces 4K scene training time from several hours to around 10 minutes by employing dilated rendering, a convergence-aware training schedule, and position-appearance-based densification, consistently maintaining or improving rendering quality metrics like LPIPS.
Key insights
- For high-resolution inputs, the computational bottleneck in 3DGS shifts from alpha-blending to earlier rasterization stages like tile-Gaussian pair generation and depth sorting.
- The convergence of 3DGS training loss often follows an approximate power-law relationship, which can be leveraged for dynamic and efficient management of the Gaussian densification budget.
- Color gradients provide valuable information for Gaussian densification in textureless regions where positional gradients are negligible, enabling more complete and detailed scene reconstruction.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.