University of Bonn Brings Geometry Processing Directly to 3D Gaussian Splatting
A new discrete Laplace-Beltrami operator uses Mahalanobis distance to run geometry processing directly on 3D Gaussian splats, skipping mesh conversion.
Takeaways
- New Laplace-Beltrami operator runs directly on 3D Gaussian splats, no mesh extraction required. Paper
- Uses Mahalanobis distance to build neighborhoods that respect each Gaussian's anisotropic covariance.
- Symmetric graph-based outlier removal prunes interior and low-opacity splats before geometry processing.
- Adaptive training integrates pruning into 3DGS optimization, shrinking scenes by roughly 10x versus reconstructed meshes.
- Beats mesh-based and point cloud Laplacians on curvature error (16.02 vs 43.05) and geodesic accuracy.
- Code and TOSCA-rendered dataset released; accepted at 3DV 2026. GitHub
3D Gaussian Splatting has taken over as the go-to representation for photorealistic scene rendering, but the moment you want to do anything geometric with it, like computing curvature, geodesics, or shape correspondence, the pipeline breaks. A new paper from researchers at the University of Bonn and the Lamarr Institute, Laplace-Beltrami Operator for Gaussian Splatting, closes that gap by defining a native geometry processing operator that runs directly on Gaussian splats without touching a mesh.
Why Gaussian splats are painful to analyze
3D Gaussian Splatting represents a scene as thousands of anisotropic 3D Gaussians, each with a mean, covariance, opacity, and color. The optimization process aims to match the render of these parameters to a collection of images from different view points which leads to bias towards photometric consistency instead of geometric accuracy, leading to Gaussians that may not be localized exactly on the object surface but still produce visually compelling results.
That bias creates two headaches for anyone trying to do geometry processing. Many Gaussians may be positioned in free space or within object interiors if they contribute to visual quality, and low-opacity Gaussians that minimally affect rendering quality are retained in the representation, creating outliers for geometric analysis. The usual workarounds, extracting a mesh with Poisson reconstruction or treating the Gaussian centers as a raw point cloud, either introduce reconstruction artifacts or throw away the covariance information that describes local surface orientation.