Research map / 3D vision / Depth and stereo
VGGT-SLAM: Dense RGB SLAM Optimized on the SL(4) Manifold
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.
Builds a monocular dense RGB SLAM system by aligning VGGT submaps via 15-DOF SL(4) homographies optimized in a factor graph with loop closures, arguing similarity transforms are insufficient for uncalibrated cameras.
Problem
Extending feed-forward multi-view reconstruction (VGGT) to long uncalibrated monocular video, where GPU memory limits the number of frames; submaps are misaligned under similarity transforms because of projective ambiguity.
Approach
Align submaps by estimating 15-DOF homographies (SL(4) group) between sequential and loop-closure submaps via RANSAC 5-point solve, and optimize a pose graph on the SL(4) manifold using Levenberg-Marquardt, rather than Sim(3) alignment.
Key results
- On TUM RGB-D, the SL(4) version of VGGT-SLAM (w=32) achieves the best average ATE of 0.053 m, versus 0.060 m for MASt3R-SLAM* (uncalibrated).
- On 7-Scenes, VGGT-SLAM (SL(4), w=32) matches MASt3R-SLAM's average ATE (0.067 vs 0.066 m) and achieves the best accuracy (0.052) and Chamfer distance (0.055) in dense reconstruction.
- SL(4) alignment is needed when disparity is small: with tau_disparity=0, Sim(3) alignment degrades and fails to align submaps in some scenes.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SL(4) instead of Sim(3) submap alignment (ablation) | VGGT-SLAM · TUM RGB-D | ATE RMSE: 0.074 -> 0.053 | Helped |
| use SL(4) instead of Sim(3) submap alignment (ablation) | VGGT-SLAM · 7-Scenes | ATE RMSE: 0.067 -> 0.067 | No clear effect |
| use SL(4) instead of Sim(3) submap alignment (ablation) | VGGT-SLAM · 7-Scenes | Chamfer RMSE: 0.057 -> 0.055 | Helped |
| use SL(4) instead of Sim(3) submap alignment (ablation) | VGGT-SLAM · 7-Scenes | Completion RMSE: 0.062 -> 0.058 | Hurt |
| use SL(4) instead of Sim(3) submap alignment (ablation) | VGGT-SLAM · TUM RGB-D | ATE RMSE on floor: 0.254 -> 0.141 | Hurt |
| add loop closures (ablation) | VGGT-SLAM · TUM RGB-D | ATE | Helped |
| increase the number of loop closures as the number of submaps increases (ablation) | VGGT-SLAM · TUM RGB-D | ATE | Helped |
| increase confidence threshold (ablation) | VGGT-SLAM · 7-Scenes | Accuracy RMSE | Helped |
| increase confidence threshold (ablation) | VGGT-SLAM · 7-Scenes | Completion RMSE | Hurt |
| reduce submap size from w=32 to w=8 (ablation) | VGGT-SLAM (Sim(3)) · TUM RGB-D | ATE RMSE: 0.074 -> 0.040 | Helped |
| reduce submap size from w=32 to w=8 (ablation) | VGGT-SLAM (SL(4)) · TUM RGB-D | ATE RMSE: 0.053 -> 0.115 | Hurt |
Surprising findings
- Sim(3) alignment, the natural baseline for VGGT submaps, fails under projective ambiguity, especially when inter-frame disparity is small.
- The planar TUM floor scene causes SL(4) estimation to diverge due to degenerate homography solutions.
Limitations
- Homography estimation is degenerate for planar scenes, causing divergence on TUM floor.
- Sensitive to outliers in VGGT depth; 5-point RANSAC can fail on pure-rotation sequences with small submaps.
- 15-DOF freedom admits perspective drift that is harder to correct than classical SLAM drift.
Experimental setup
Datasets: 7-Scenes TUM RGB-D Clio
Benchmarks: 7-Scenes TUM RGB-D
Metrics: ATE RMSE accuracy completion Chamfer distance
Baselines: DROID-SLAM MASt3R-SLAM Spann3R DPV-SLAM GO-SLAM ORB-SLAM3 NICER-SLAM
Builds on: VGGT MASt3R-SLAM SALAD DUSt3R
Closest papers by meaning
- VTGaussian-SLAM: RGBD SLAM for Large Scale Scenes with Splatting View-Tied 3D Gaussians (ICML 2025 · similarity 0.89)
- SLAM3R: Real-Time Dense Scene Reconstruction from Monocular RGB Videos (CVPR 2025 · similarity 0.88)
- VGGT: Visual Geometry Grounded Transformer (CVPR 2025 · similarity 0.88)
- WildGS-SLAM: Monocular Gaussian Splatting SLAM in Dynamic Environments (CVPR 2025 · similarity 0.88)
- Gaussian Splatting SLAM (CVPR 2024 · similarity 0.87)
- Photo-SLAM: Real-time Simultaneous Localization and Photorealistic Mapping for Monocular Stereo and RGB-D Cameras (CVPR 2024 · 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.