Reproduce “CaT-GS: Efficient 3DGS Rendering for Large Scale Scenes via Inter-frame Caching and Tile Scheduling” 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
Tingjia Zhang, Bo Chen, Shengzhong Liu, Fan Wu, Guihai Chen
CaT-GS averages 241.5 FPS on UAV-1, versus 129.2 for Flash-GS and 25.2 for original 3DGS, using a 6,895,821-Gaussian model. It predicts camera-motion trails to reuse visibility and sorting lists across later frames, and splits overloaded tile tasks. On video-captured UAV-1 frames, PSNR/SSIM are 30.15/0.935 for 3DGS, 30.09/0.931 for key frames, and 30.05/0.926 for sub-frames.
Key insights
- Significant performance bottlenecks in 3DGS for large scenes arise from inter-frame redundancies in pre-processing and imbalanced GPU tile workloads, issues not fully addressed by prior model compression or intra-frame optimization methods.
- Speculatively pre-calculating Gaussian visibility and sorting for a small batch of future frames can effectively exploit viewpoint coherence, allowing sub-frames to reuse cached data and skip costly pipeline stages.
- Dynamic load balancing by splitting "heavy" tiles into smaller tasks, combined with refactored alpha blending, is vital for improving GPU utilization and accelerating the rasterization phase, even if it leads to processing slightly more Gaussians overall.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.