Reproduce “SpaRV: Sparse Radiance Voxels for Scalable Geometrically Consistent LiDAR-Visual Mapping and Rendering” 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

Jianheng Liu, Yunfei Wan, Fangming Cheng, Xiangyu Hong, Zuhao Zou, Haotian Li, Longji Yin, Chunran Zheng and 2 more

SpaRV improves robot maps so they render more reliably from viewpoints beyond the route. It extends SVRaster’s sparse-voxel representation with differentiable ray marching that jointly uses camera rays and irregular LiDAR measurements, plus geometry guidance. Across eight Replica scenes, it achieves 36.690 dB average extrapolated PSNR versus 33.347 dB for SVRaster; this measures rendered-image fidelity, not geometric error.

Key insights

  • Sparse LiDAR depth alone does not constrain camera-visible regions between measured rays; occupied-voxel guidance uses observed surfaces to discourage floating geometry in front of them.
  • Refining camera parameters against the evolving map can improve alignment: on FAST-LIVO2 scenes, SpaRV’s average training-view PSNR rises from 26.55 to 29.87 dB.
  • A shared ray-marching representation can support ordinary and non-pinhole cameras, irregular LiDAR patterns, and level-of-detail rendering within a bounded GPU budget.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.