Reproduce “Proxy-GS: Unified Occlusion Priors for Training and Inference in Structured 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
Yuanyuan Gao, Yuning Gong, Yifei Liu, Li Jingfeng, Dingwen Zhang, Yanci Zhang, Dan Xu, Xiao Sun and 1 more
Proxy-GS uses a simplified mesh to guide visibility in large-scene 3D rendering. Its depth maps cull hidden anchors during rendering and guide new anchors toward visible surfaces during training; 1000×1000 maps take under 1 ms. On MatrixCity Block 5, it reaches 151 FPS and 21.68 PSNR, versus Octree-GS at 48 FPS and 21.41; gains vary by scene and require a proxy mesh.
Key insights
- Explicitly integrating occlusion awareness into both training and inference of MLP-based 3DGS significantly reduces computational redundancy and boosts rendering efficiency, especially in complex, occlusion-rich scenes.
- Leveraging lightweight geometric proxies with hardware rasterization provides an extremely fast and GPU-resident mechanism for accurate occlusion culling, making real-time performance gains practical.
- Guiding the densification of Gaussians towards visible surfaces during training, rather than purely error-driven methods, leads to more geometrically consistent representations and prevents the generation of useless primitives in occluded space.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.