Direct Mesh Extraction from 3D Gaussian Splatting

Develop a method to extract high-quality meshes or other explicit 3D geometries directly from 3D Gaussian Splatting representations without converting to alternative representations such as InstantNGP, enabling reliable geometry extraction within the Gaussian Splatting framework itself.

Background

The paper adopts a two-stage pipeline: Stage I performs interactive generation using 3D Gaussian Splatting to enable direct user manipulation, while Stage II converts the Gaussian representation into InstantNGP for geometry refinement and mesh extraction. The authors note that Gaussian Splatting is well-suited for interaction but struggles with high-quality geometry formation under SDS supervision and lacks a direct pathway to extract explicit geometry.

Because mesh extraction from Gaussian Splatting is not currently available, the method relies on converting the Gaussian representation to InstantNGP via NeRF distillation to enable mesh extraction and further refinement. This highlights an unresolved problem: creating a reliable technique for direct mesh (or other explicit geometry) extraction from Gaussian Splatting without representation conversion.