Reproduce “A LoD of Gaussians: Unified Training and Rendering for Ultra-Large Scale Reconstruction with External Memory” 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
Felix Windisch, Thomas Köhler, Lukas Radl, Mattia D'Urso, Michael Steiner, Dieter Schmalstieg, Markus Steinberger
A hierarchical, out-of-core 3D Gaussian Splatting system enables coherent reconstruction and interactive rendering of ultra-large, mixed-scale scenes on a single consumer GPU. By combining view-dependent level-of-detail selection, dynamic hierarchy maintenance, CPU-memory storage, and cache-aware streaming, the method supports models with more than 150 million Gaussians while avoiding many chunk-boundary artifacts; caching reduces RAM-to-GPU transfers by up to 93% and approximately doubles rendering speed in reported tests.
Key insights
- Avoiding spatial partitioning allows one coherent Gaussian model to represent views that cross arbitrary scene regions, reducing the boundary artifacts associated with independently trained chunks.
- The hierarchy is part of the optimization process: parent nodes preserve coarse structure while densification adds children where screen-space gradients indicate missing detail.
- Cache reuse is central to practical out-of-core performance; scheduling nearby camera views and caching SPT cuts greatly lowers the number of Gaussians loaded from host memory.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.