Reproduce “TopoSurfel: Closing the Loop between Gaussian Surfels and Meshes for Surface 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
Chuanjin Fan, Wenjie Chang, Bohao Liao, Yujia Chen, Wenfei Yang, Tianzhu Zhang
TopoSurfel reconstructs usable surfaces from multi-view images by coupling flat Gaussian surfels with a temporary connected mesh during training. A warm-up initializes surfels, while differentiable Poisson reconstruction and marching cubes provide geometry feedback for normals, depth, and density control. On six Tanks and Temples scenes, it reaches mean F1 0.52 versus MILo’s 0.49, but not every scene improves.
Key insights
- A continuous proxy mesh can be differentiably extracted from discrete Gaussian surfels mid-optimization and effectively used as an explicit global prior to refine surface geometry.
- Direct geometric feedback, such as aligning surfel normals with mesh normals and using mesh distance for pruning and densification, is crucial for producing high-fidelity, coherent surfaces and suppressing artifacts.
- Combining an initial warm-up with coarse TSDF re-initialization and a hybrid strategy for handling foreground/background regions significantly improves reconstruction robustness in complex and large-scale scenes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.