Research map / 3D vision / Gaussian splatting
SuGaR: Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SuGaR, a regularization term that aligns 3D Gaussians with scene surfaces, followed by fast Poisson-based mesh extraction and an optional joint refinement that binds Gaussians to the mesh for high-quality mesh rendering.
Problem
Extracting accurate, editable meshes from unstructured millions of 3D Gaussians produced by 3D Gaussian Splatting, which Marching Cubes fails on, while keeping high-quality rendering.
Approach
A surface-alignment regularizer compares a density-derived SDF with the SDF from Gaussian depth maps, so Gaussians lie flat on the surface; meshes are then extracted by Poisson reconstruction from points sampled on a density level set, and Gaussians bound to mesh triangles are jointly refined for rendering.
Key results
- R-SuGaR-15K reaches 27.27 average PSNR on Mip-NeRF360 versus 23.15 for NeRFMeshing and 24.87 for R-SuGaR-2K among mesh-based methods.
- Mesh extraction on Mip-NeRF360 with the proposed level-set Poisson method gives PSNR 24.87 versus 23.91 for Marching Cubes.
- Mesh extraction takes 5-10 minutes and optimization 15-45 minutes on a single V100 GPU, versus at least 24 hours for neural-SDF methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SuGaR mesh extraction and Gaussian-bound mesh rendering | SuGaR · Mip-NeRF360, average on all scenes | PSNR: 23.15 -> 27.27 | Helped |
| use SuGaR mesh extraction and Gaussian-bound mesh rendering | SuGaR · Mip-NeRF360, average on all scenes | PSNR: 28.69 -> 27.27 | Hurt |
| increase refinement iterations from 2,000 to 7,000 (ablation) | SuGaR · Mip-NeRF360, average on all scenes | PSNR: 24.87 -> 26.77 | Helped |
| increase refinement iterations from 7,000 to 15,000 (ablation) | SuGaR · Mip-NeRF360, average on all scenes | PSNR: 26.77 -> 27.27 | Helped |
| replace Marching Cubes with SuGaR's level-set point sampling and Poisson reconstruction (ablation) | SuGaR · Mip-NeRF360 | PSNR: 23.91 -> 24.87 | Helped |
| use level-set points instead of Gaussian centers for Poisson reconstruction (ablation) | SuGaR · Mip-NeRF360 | PSNR: 23.76 -> 24.87 | Helped |
| change the extracted density level from 0.1 to 0.3 (ablation) | SuGaR · Mip-NeRF360 | PSNR: 24.62 -> 24.87 | Helped |
| change the extracted density level from 0.3 to 0.5 (ablation) | SuGaR · Mip-NeRF360 | LPIPS: 0.304 -> 0.304 | No clear effect |
| use surface-aligned Gaussians instead of a UV texture (ablation) | SuGaR mesh rendering · Mip-NeRF360, 1M vertices | LPIPS: 0.478 -> 0.295 | Helped |
| use surface-aligned Gaussians instead of a UV texture (ablation) | SuGaR mesh rendering · Mip-NeRF360, 200K vertices | LPIPS: 0.419 -> 0.300 | Helped |
| increase mesh resolution from 200K to 1M vertices with surface-aligned Gaussians (ablation) | SuGaR mesh rendering · Mip-NeRF360 | PSNR: 24.24 -> 24.51 | Helped |
Surprising findings
- Gaussians bound to the mesh outperform optimized UV textures on the same meshes (24.51 vs 21.24 PSNR at 1M vertices).
Limitations
- Gaussians can cheat geometry with cavities to reproduce specular effects, distorting surfaces in intensely specular regions.
- The surface assumption complicates rendering volumetric effects or fuzzy materials.
- Mip-NeRF360 evaluation uses 7 of 9 scenes due to unavailable scenes.
Experimental setup
Datasets: Mip-NeRF360 DeepBlending Tanks&Temples
Benchmarks: Mip-NeRF360 Tanks&Temples DeepBlending
Metrics: PSNR SSIM LPIPS
Baselines: 3D Gaussian Splatting Marching Cubes Poisson (centers) Mobile-NeRF NeRFMeshing BakedSDF Mip-NeRF360 Plenoxels Instant-NGP
Builds on: 3D Gaussian Splatting Poisson reconstruction Marching Cubes quadric error simplification
Closest papers by meaning
- Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 2025 · similarity 0.92)
- GaussianPro: 3D Gaussian Splatting with Progressive Propagation (ICML 2024 · similarity 0.91)
- Mani-GS: Gaussian Splatting Manipulation with Triangular Mesh (CVPR 2025 · similarity 0.91)
- GaussianSpa: An "Optimizing-Sparsifying" Simplification Framework for Compact and High-Quality 3D Gaussian Splatting (CVPR 2025 · similarity 0.91)
- VA-GS: Enhancing the Geometric Representation of Gaussian Splatting via View Alignment (NeurIPS 2025 · similarity 0.91)
- Lightweight Predictive 3D Gaussian Splats (ICLR 2025 · similarity 0.91)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.