Paper
Khoi Do
Yu-kun Lai
Binh-Son Hua
Recent advances in 3D Gaussian Splatting (3DGS) have demonstrated its effectiveness for photorealistic novel view synthesis. However, extracting clean, watertight, explicit surface meshes from 3DGS remains a critical challenge. Previous methods approximate geometry via discrete Marching Tetrahedra, where each grid vertex is assigned a single scalar signed distance value. At complex geometric boundaries and thin structures, this suboptimal parameterization forces severe compromises across shared vertices, causing gradient collisions and geometric frustration. Furthermore, existing optimization objectives neglect element shape conditioning, producing degenerate triangles. We introduce TopSDF, a framework that models the signed distance state as a multivariate Gaussian random field. By learning latent topological covariances across tetrahedral vertices, our formulation expands local degrees of freedom for SDF estimates, allowing the candidate mesh vertices to be placed flexibly along the bipolar edges in the tetrahedra as driven by such topological correlation. Concurrently, we directly integrate variational meshing energies into the joint optimization, coupled with an energy-driven Gaussian densification strategy to improve triangle mesh quality. Our experiments across Mip-NeRF 360, Tanks & Temples, and ScanNet++ show that TopSDF achieves state-of-the-art mesh-based novel view rendering (+0.86 dB PSNR on Mip-NeRF 360) and triangle-mesh quality (from 0.37 to 0.49 aspect ratio).
PDF URL