Reproduce “HiGS: A Hierarchical Rendering Architecture for Real-Time 3D Gaussian Splatting” 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
Dawid Pająk, Martin Bisson, Rodolfo Lima
HiGS renders captured 3D scenes from new views by sorting coarse macro-tiles while rasterizing their fine tiles, splitting Gaussian lists into bounded batches and dynamically assigning dense work. Across seven Mip-NeRF 360 scenes, it reaches mean 4K throughput of 1,214 FPS versus 670 for FlashGS, or 1.81×, on specified hardware and settings. Quality remains similar, not bitwise identical.
Key insights
- Decoupling the granularities of spatial partitioning and rasterization via a hierarchical structure effectively resolves the fundamental efficiency trade-off in 3DGS, allowing each stage to operate optimally.
- Density-proportional work decomposition in rasterization dynamically balances load across GPU units, effectively mitigating resolution-dependent bottlenecks and 'rasterizer tail effects'.
- Asynchronous segmented depth sorting dramatically reduces sorting overhead by processing independent macro-tile segments, enabling efficient on-device sorting without host intervention.
- Half-precision arithmetic can be successfully integrated into a high-fidelity rendering pipeline to reduce memory bandwidth and increase throughput, provided numerical stability challenges are systematically addressed.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.