TopoSurfel visual showing Gaussian surfels closing the loop with a proxy mesh

TopoSurfel: Closing the Loop between Gaussian Surfels and Meshes

TopoSurfel is a lightweight surface reconstruction framework in which discrete Gaussian surfels and a continuous proxy mesh co-evolve, producing cleaner and more complete geometry without additional learnable parameters.

Chuanjin Fan·Wenjie Chang·Bohao Liao·Yujia Chen·Wenfei Yang·Tianzhu Zhang

University of Science and Technology of China

VIDEO

Project Video

Watch the full project video directly on this page.

CORE IDEA

Bring the mesh into the loop.

Instead of extracting a mesh only after training, TopoSurfel continuously uses a differentiable proxy mesh as structured geometry guidance.

01

Stable warm-up

Initialize Gaussian surfels and an explicit surface prior from sparse multi-view observations.

02

Proxy mesh

Extract a continuous differentiable mesh directly from oriented surfels, without auxiliary networks.

03

Closed-loop evolution

Mesh-guided normals and density control suppress floaters, fill holes, and refine the next mesh.

METHOD

Surfel–mesh co-evolution

Warm up once, then iteratively exchange geometry between the two representations.

TopoSurfel pipeline with warm-up and closed-loop optimization stages
GEOMETRY RESULTS

Interactive DTU reconstruction

Drag to rotate · Scroll to zoom · Double-click to reset

All meshes use a neutral material and a shared display budget for a geometry-only comparison.
DTU scan
2DGS
GOF
MILo
TopoSurfel Ours
0.51DTU Chamfer Distance ↓
37 minDTU training time
0Extra learnable parameters
284 FPSReal-time rendering
CITATION

Cite TopoSurfel.

If you find this work useful, please use the following BibTeX entry.

BibTeX
@article{fan2026toposurfel,
  title   = {TopoSurfel: Closing the Loop between Gaussian
             Surfels and Meshes for Surface Reconstruction},
  author  = {Fan, Chuanjin and Chang, Wenjie and Liao, Bohao
             and Chen, Yujia and Yang, Wenfei and Zhang, Tianzhu},
  journal = {arXiv preprint arXiv:2608.20687},
  year    = {2026},
  url     = {https://arxiv.org/abs/2608.20687}
}