Reproduce “VastGaussian: Vast 3D Gaussians for Large Scene Reconstruction” 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
Jiaqi Lin, Zhihao Li, Xiao Tang, Jianzhuang Liu, Shiyong Liu, Jiayue Liu, Yangdi Lu, Xiaofei Wu and 3 more
VastGaussian extends 3D Gaussian Splatting to large-scale scenes, offering high-fidelity reconstruction and real-time rendering for extensive environments. It reduces training time for large scenes to approximately 2.5 hours from a previous range of 19-42 hours while maintaining rendering speeds of 126-171 FPS at 1080p.
Key insights
- An airspace-aware visibility criterion for camera selection within partitioned cells is critical for supervising empty space and preventing "floater" artifacts.
- Decoupling appearance corrections from the core 3D Gaussian representation enables robust handling of varied lighting in training data without impacting real-time inference.
- Parallel optimization of scene cells, combined with careful data selection, drastically reduces training time and GPU memory requirements for large-scale 3DGS.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.