Research map / 3D vision / Gaussian splatting
AtlasGS: Atlanta-world Guided Surface Reconstruction with Implicit Structured Gaussians
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 AtlasGS, an Atlanta-world guided implicit-structured Gaussian Splatting that adds semantic Gaussians and plane-indicator regularization to reconstruct smooth indoor and urban surfaces efficiently.
Problem
Surface reconstruction of indoor and urban scenes suffers from low-texture regions lacking global geometric consistency, and Gaussian Splatting and implicit SDF methods trade off smoothness, detail, and efficiency.
Approach
Embeds a sparse voxel feature grid with shared MLP decoders into 2DGS-style surfels so neighboring Gaussians are coupled, adds semantic Gaussian lifting (wall/floor/ceiling/other), and regularizes Gaussians with learnable explicit gravity-aligned plane indicators under the Atlanta-world assumption.
Key results
- On ScanNet, the method reaches 3.62 Acc, 3.93 Comp and 77.98 F-score, versus 2DGS at 39.27 F-score, while training in 27 minutes.
- On MatrixCity, CD is 0.028 versus 0.091 for the next-best GaussianPro (second-best listed).
- On Replica, the F-score is 87.35 compared with 73.08 for MonoSDF and 70.37 for GSRec; novel view synthesis PSNR on Replica (39.58) is below 2DGS (41.59).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AtlasGS instead of the strongest listed baseline | AtlasGS · Replica | F-score: 70.37 -> 87.35 | Helped |
| Use AtlasGS instead of the strongest listed baseline | AtlasGS · ScanNet++ | F-score: 76.67 -> 87.48 | Helped |
| Use AtlasGS instead of the strongest listed baseline | AtlasGS · ScanNet | F-score: 71.21 -> 77.98 | Helped |
| Use AtlasGS instead of the strongest listed baseline | AtlasGS · MatrixCity | CD: 0.091 -> 0.028 | Helped |
| Use AtlasGS instead of 2DGS | AtlasGS · Replica | PSNR: 41.59 -> 39.58 | Hurt |
| Use AtlasGS instead of GSRec | AtlasGS · ScanNet++ | SSIM: 0.8314 -> 0.8321 | Helped |
| Add implicit-structured Gaussians to 2DGS with normal and depth priors (ablation) | AtlasGS · ScanNet | geometry quality (CD and F-score) | Helped |
| Add 3D global planar regularization (ablation) | AtlasGS · ScanNet | geometry quality | Helped |
| Add 2D local surface regularization (ablation) | AtlasGS · ScanNet | geometry quality | Helped |
| Add the normal prior (ablation) | AtlasGS · ScanNet | geometry quality | Helped |
| Add the depth prior (ablation) | AtlasGS · ScanNet | geometry quality | Helped |
Surprising findings
- Rendering is slower (70 FPS on ScanNet) than other Gaussian baselines such as Scaffold-GS (279 FPS) and 2DGS (118 FPS), as the method decodes all Gaussians when rendering.
- 2DGS achieves higher PSNR and SSIM than the proposed method on Replica novel view synthesis.
Limitations
- Training and rendering are slower than previous Gaussian-based methods.
- Depends on a pretrained semantic segmentation model for a limited set of structural classes (wall, floor, ceiling).
- Atlanta-world assumption limits applicability beyond man-made scenes with dominant horizontal and vertical planes.
Experimental setup
Datasets: ScanNet ScanNet++ Replica MatrixCity
Benchmarks: ScanNet ScanNet++ Replica MatrixCity
Metrics: Accuracy Completeness Chamfer Distance Precision Recall F-score PSNR SSIM LPIPS FPS
Baselines: ManhattanSDF MonoSDF Scaffold-GS 2DGS DN-Splatter GSRec GaussianPro
Builds on: 2D Gaussian Splatting Scaffold-GS Atlanta world assumption MonoSDF RANSAC
Closest papers by meaning
- IndoorGS: Geometric Cues Guided Gaussian Splatting for Indoor Scene Reconstruction (CVPR 2025 · similarity 0.91)
- PlanarGS: High-Fidelity Indoor 3D Gaussian Splatting Guided by Vision-Language Planar Priors (NeurIPS 2025 · similarity 0.91)
- Horizon-GS: Unified 3D Gaussian Splatting for Large-Scale Aerial-to-Ground Scenes (CVPR 2025 · similarity 0.90)
- Evolving High-Quality Rendering and Reconstruction in a Unified Framework with Contribution-Adaptive Regularization (CVPR 2025 · similarity 0.89)
- EAP-GS: Efficient Augmentation of Pointcloud for 3D Gaussian Splatting in Few-shot Scene Reconstruction (CVPR 2025 · similarity 0.89)
- SceneSplat++: A Large Dataset and Comprehensive Benchmark for Language Gaussian Splatting (NeurIPS 2025 · similarity 0.89)
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.