Research map / 3D vision / Gaussian splatting
CoMapGS: Covisibility Map-based Gaussian Splatting for Sparse Novel View Synthesis
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 CoMapGS, which builds pixel-wise covisibility maps from MASt3R correspondences to enhance sparse COLMAP point clouds and applies covisibility-weighted proximity-classifier supervision to improve sparse-view 3D Gaussian Splatting.
Problem
Sparse-view novel view synthesis with 3DGS suffers from region-wise imbalanced supervision: frequently covisible regions are well reconstructed while mono-view (high-uncertainty) regions are under-supervised, and sparse COLMAP initial point clouds lack geometric detail.
Approach
Use per-pixel covisibility counts (from dense correspondences) to (1) densify initial point clouds with triangulated MASt3R points and monocular-depth points rescaled by a learned linear scaling, and (2) weight a proximity loss from an MLP classifier so that mono-view Gaussians receive stronger supervision, extended outside the frustum for high-covisibility scenes.
Key results
- On LLFF 3-view, CoR-GS + CoMapGS reaches 21.105 PSNR / 0.747 SSIM / 0.182 LPIPS vs CoR-GS reproduced 20.473 / 0.717 / 0.199.
- On Mip-NeRF 360 12-view, CoR-GS + CoMapGS achieves 19.680 PSNR and 0.394 LPIPS vs CoR-GS 19.162 and 0.414.
- Ablation on LLFF 6-view: full method raises CoR-GS PSNR from 24.777 to 25.204 and LPIPS from 0.116 to 0.108.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add CoMapGS to CoR-GS | CoR-GS · LLFF 3-view | PSNR: 20.473 -> 21.105 | Helped |
| Add CoMapGS to CoR-GS | CoR-GS · LLFF 6-view | PSNR: 24.777 -> 25.204 | Helped |
| Add CoMapGS to CoR-GS | CoR-GS · LLFF 9-view | PSNR: 26.475 -> 26.731 | Helped |
| Add CoMapGS to CoR-GS | CoR-GS · Mip-NeRF 360 12-view | LPIPS: 0.414 -> 0.394 | Helped |
| Add CoMapGS to CoR-GS | CoR-GS · Mip-NeRF 360 24-view | PSNR: 23.321 -> 23.462 | Helped |
| Add proximity-loss supervision without enhancing the initial point cloud (ablation) | CoR-GS · LLFF 6-view | PSNR: 24.777 -> 24.787 | No clear effect |
| Enhance initial points only in covisible regions (ablation) | CoR-GS · LLFF 6-view | PSNR: 24.777 -> 24.90 | Helped |
| Add proximity loss to covisible-region point enhancement (ablation) | CoR-GS · LLFF 6-view | PSNR: 24.90 -> 25.153 | Helped |
| Enhance initial points in both covisible and mono-view regions (ablation) | CoR-GS · LLFF 6-view | PSNR: 24.777 -> 25.076 | Helped |
| Add proximity-loss supervision to full point-cloud enhancement (ablation) | CoR-GS · LLFF 6-view | PSNR: 25.076 -> 25.204 | Helped |
Surprising findings
- ReconFusion obtains higher 3-view PSNR than CoMapGS, which the authors attribute to diffusion models optimizing for seamless colors rather than original mono-view patterns.
- Starting from sparse COLMAP point clouds without enhancement gives only limited gains from weighted supervision; enhanced initialization is needed for benefits to appear.
Limitations
- Gains in PSNR are modest on Mip-NeRF 360 at 24 views due to unseen regions in test views.
- Relies on multiple external modules (MASt3R, monocular depth, COLMAP), adding pipeline complexity.
- CoR-GS + CoMapGS is slightly below reproduced FSGS-based result or baseline in some settings (e.g., 9-view PSNR 26.292 vs CoR-GS 26.475).
Experimental setup
Datasets: LLFF Mip-NeRF 360
Benchmarks: LLFF Mip-NeRF 360
Metrics: PSNR SSIM LPIPS
Baselines: FSGS CoR-GS 3DGS ReconFusion DNGaussian FreeNeRF
Builds on: 3D Gaussian Splatting CoR-GS FSGS DyCheck
Closest papers by meaning
- Quantifying and Alleviating Co-Adaptation in Sparse-View 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.91)
- Binocular-Guided 3D Gaussian Splatting with View Consistency for Sparse View Synthesis (NeurIPS 2024 · similarity 0.89)
- FewViewGS: Gaussian Splatting with Few View Matching and Multi-stage Training (NeurIPS 2024 · similarity 0.89)
- NexusGS: Sparse View Synthesis with Epipolar Depth Priors in 3D Gaussian Splatting (CVPR 2025 · similarity 0.89)
- Generative Sparse-View Gaussian Splatting (CVPR 2025 · similarity 0.89)
- CoCoGaussian: Leveraging Circle of Confusion for Gaussian Splatting from Defocused Images (CVPR 2025 · similarity 0.88)
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.