SuGaR: Surface-Aligned Gaussian Splatting
Bridges the gap between Gaussian point clouds and traditional polygon mesh pipelines via surface-alignment regularization.
Authors / Team
Antoine Guédon · Researcher
Location
法国 · 巴黎
Year
2023
Deep Dive
SuGaR addresses a major pain point: vanilla 3DGS is hard to convert into editable surface meshes for traditional 3D software. During optimization, Gaussians usually exhibit chaotic semi-transparent overlap. SuGaR adds a regularization term forcing the 3D Gaussians to flatten and adhere tightly to the underlying object surface, after which clean polygonal meshes can be extracted via classic algorithms like Poisson reconstruction. The method significantly broadens downstream usage — letting 3DGS results flow into existing game engines and animation workflows for collision and UV editing.
What we learn
- 01
Aligning geometric normals is a prerequisite for high-quality surface extraction.
- 02
Enforcing physical constraints inside the optimization pipeline suppresses internal chaos caused by pure visual overfitting.
- 03
Converting Gaussians back to standard meshes immediately unlocks decades of classic computer graphics tooling.
Verbatim quote
"We propose a method that extracts precise and appealing meshes from 3D Gaussian Splatting."— source ↗
Tags
Links
Sources