Research map / 3D vision / Gaussian splatting
MAtCha Gaussians: Atlas of Charts for High-Quality Geometry and Photorealism From Sparse Views
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 MAtCha Gaussians, which model scene surfaces as an atlas of depth-map charts deformed by a lightweight neural model and rendered with 2D Gaussian surfels, to get sharp meshes and photorealistic novel views from sparse unposed images within minutes.
Problem
Recovering accurate, sharp explicit 3D surface meshes of unbounded scenes and photorealistic novel views from very few input images, where monocular depths have scale ambiguity and volumetric/Gaussian methods need dense views.
Approach
Represent the surface as an atlas of charts initialized from monocular depth maps, align them to SfM points with a neural deformation model (low-res chart feature grids plus depth encodings), preserve high-frequency detail with a normal/curvature structure loss, and refine with 2D Gaussian surfels attached to the charts.
Key results
- On DTU with 3 views, mean Chamfer Distance 1.04 mm vs 1.36 mm for previous best Spurfies (about 24% better).
- On Tanks&Temples with 10 views, F-Score 0.156 vs 0.144 for GOF+MASt3R-SfM and 0.121 for 2DGS+MASt3R-SfM.
- On Mip-NeRF 360 with 10 views, average PSNR 25.10 vs 24.50 for the best baseline (GOF+MASt3R-SfM); the 10% quantile PSNR improves from 20.99 to 21.55.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use MAtCha Gaussians instead of Spurfies | Sparse-view surface reconstruction · DTU (3 views) | Chamfer Distance: 1.36 -> 1.04 | Helped |
| use MAtCha Gaussians instead of the strongest reported baseline | Sparse-view surface reconstruction · Tanks&Temples (5 training images) | F-Score: 0.052 -> 0.072 | Helped |
| use MAtCha Gaussians instead of the strongest reported baseline | Sparse-view surface reconstruction · Tanks&Temples (10 training images) | F-Score: 0.144 -> 0.156 | Helped |
| add chart encodings (ablation) | MAtCha Gaussians · DTU | Chamfer Distance: 2.693 -> 1.04 | Helped |
| add chart encodings (ablation) | MAtCha Gaussians · Mip-NeRF 360 | PSNR: 16.37 -> 17.59 | Helped |
| add depth encodings (ablation) | MAtCha Gaussians · DTU | Chamfer Distance: 1.601 -> 1.04 | Helped |
| add depth encodings (ablation) | MAtCha Gaussians · Mip-NeRF 360 | PSNR: 17.38 -> 17.59 | Helped |
| add Lfit confidence (ablation) | MAtCha Gaussians · DTU | Chamfer Distance: 1.703 -> 1.04 | Helped |
| add Lfit confidence (ablation) | MAtCha Gaussians · Mip-NeRF 360 | PSNR: 17.39 -> 17.59 | Helped |
| add Lstruct (ablation) | MAtCha Gaussians · DTU | Chamfer Distance: 1.716 -> 1.04 | Helped |
| add Lstruct (ablation) | MAtCha Gaussians · Mip-NeRF 360 | PSNR: 17.00 -> 17.59 | Helped |
| add Lalign (ablation) | MAtCha Gaussians · DTU | Chamfer Distance: 1.565 -> 1.04 | Helped |
Surprising findings
- On Tanks&Temples with 5 views, the 2DGS+MASt3R-SfM baseline (F-Score 0.052) and GOF+MASt3R-SfM (0.054) beat MAtCha (0.072) only in the sense that MAtCha is higher; on 10 views MAtCha is best on F-Score, while on 5 views it is the best too, so no reversal was found in this table.
- Replacing charts with unadjusted or coarse MASt3R-SfM depths makes Gaussian regularization detrimental in sparse views, indicating the charts are essential.
Limitations
- Significant lack of overlap between views remains challenging.
- Failures when both monodepth and SfM are unreliable, e.g., strong specularities or transparency.
- 2DGS rasterization requires centered principal points, so input images were cropped, leading to incomplete meshes on some DTU scans.
Experimental setup
Datasets: DTU Tanks and Temples Mip-NeRF 360 DeepBlending
Benchmarks: DTU Tanks and Temples Mip-NeRF 360 DeepBlending
Metrics: Chamfer Distance F-Score PSNR 10% quantile PSNR SSIM
Baselines: Points2Surf CAP-UDF NeuS VolSDF SuGaR 2D Gaussian Splatting + MASt3R-SfM Gaussian Opacity Fields + MASt3R-SfM SparseNeus VolRecon S-VolSDF NeuSurf Spurfies
Builds on: 2D Gaussian Splatting Gaussian Opacity Fields Depth Anything V2 MASt3R-SfM AtlasNet
Closest papers by meaning
- 3D Gaussian Flats: Hybrid 2D/3D Photometric Scene Reconstruction (NeurIPS 2025 · similarity 0.88)
- SurfelSplat: Learning Efficient and Generalizable Gaussian Surfel Representations for Sparse-View Surface Reconstruction (NeurIPS 2025 · similarity 0.88)
- DySurface: Consistent 4D Surface Reconstruction via Bridging Explicit Gaussians and Implicit Functions (NeurIPS 2026 · similarity 0.88)
- Geometry Field Splatting with Gaussian Surfels (CVPR 2025 · similarity 0.88)
- Neural Signed Distance Function Inference through Splatting 3D Gaussians Pulled on Zero-Level Set (NeurIPS 2024 · similarity 0.88)
- GeoSVR: Taming Sparse Voxels for Geometrically Accurate Surface Reconstruction (NeurIPS 2025 · similarity 0.87)
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.