Analyzing Gaussian Frosting: A mesh representation for 3D Gaussian Splatting

With: Félix Fourreau
Ressources: Report; Slides
When: 2025
Associated to: MVA

Project Description

The project was based on the Gaussian Frosting paper, by Antoine Guédon and Vincent Lepetit. Their method extracts accurate and editable meshes from 3D Gaussian Splatting representations within minutes on a single GPU. The specifity of their meshes is that contain gaussian splats, which allows volumetric rendering and thus better preserves the quality of the 3D GS representation.

After understanding the details of how 3DGS and SuGaR works, we delved into the methodology of Gaussian Frosting, and used this method ourselves to animate a figurine and render fuzzy scenes.

This project also includes a complete related work section, were take a deep dive into image based rendering (IBR) methods, their evolution, and the current landcape of the research domain.

Abstract

Recently, 3D gaussian splatting demonstrated promising results in inverse image rendering by representing radiance fields as a set of gaussians, enabling fast and accurate 3D reconstruction. However, gaussian representations are incompatible with mesh-based 3D editing software. The SuGaR method addresses this by retrieving a mesh from the gaussians, facilitating integration with standard 3D modeling tools for editing, animating, sculpting and relightning. Although SuGaR is significantly faster than state-of-the-art Neural Signed Distance Fields (neural SDF), the conversion to a mesh incurs visual quality loss as the gaussians are constrained to the mesh, which limits volumetric rendering. To overcome this, Gaussian Frosting creates a ”frosting” layer that encases the mesh, used for volumetric rendering using gaussian splatting. Additionally, the proposed approach supports deformable meshes, automatically adjusting the radiance field carried by the mesh, therefore enabling easy scene editing while maintaining competitive visual quality.