Research map / 3D vision / Depth and stereo
MASt3R-SLAM: Real-Time Dense SLAM with 3D Reconstruction Priors
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 real-time monocular dense SLAM system around the off-the-shelf two-view MASt3R 3D reconstruction prior, using efficient iterative projective pointmap matching, ray-based tracking with pointmap fusion, retrieval-based loop closure, and second-order global optimisation.
Problem
Real-time, plug-and-play dense monocular SLAM that yields globally consistent poses and dense geometry from in-the-wild videos without known or fixed camera models.
Approach
Uses MASt3R two-view pointmaps as the unifying prior for tracking, mapping and relocalisation, assuming only a generic central camera; introduces iterative projective (ray-based) pointmap matching, ray-error tracking with running weighted pointmap fusion, incremental ASMK loop closure, and a Sim(3) Gauss-Newton backend.
Key results
- On TUM RGB-D with calibration, the system reaches average ATE 0.030 m, beating DROID-SLAM (0.038 m) and GO-SLAM (0.035 m).
- On ETH3D-SLAM train, it achieves the best mean ATE (0.086 m) and AUC (23.935) among compared monocular methods.
- Runs at roughly 15 FPS on an RTX 4090, with the matching kernel taking 2 ms versus 2000 ms for dense MASt3R matching.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use calibrated MASt3R-SLAM instead of GO-SLAM | monocular SLAM · TUM RGB-D calibrated | ATE (m): 0.035 -> 0.030 | Helped |
| use uncalibrated MASt3R-SLAM instead of DROID-SLAM* | monocular SLAM · TUM RGB-D uncalibrated | ATE (m): 0.158 -> 0.060 | Helped |
| use Ours instead of DROID-SLAM | monocular SLAM · 7-Scenes | ATE (m): 0.049 -> 0.047 | Helped |
| add feature refinement to projective matching (ablation) | MASt3R-SLAM matching · TUM / 7-Scenes / EuRoC | ATE [m] w/ calib: 0.062 -> 0.039 | Helped |
| add feature refinement to projective matching (ablation) | MASt3R-SLAM matching · matching runtime | Matching Time [ms]: 0.5 -> 2 | Helped |
| use weighted-average pointmap fusion instead of the most recent pointmap (ablation) | MASt3R-SLAM · TUM / 7-Scenes / EuRoC without calibration | ATE [m]: 0.207 -> 0.097 | Helped |
| use weighted-average pointmap fusion instead of the first pointmap (ablation) | MASt3R-SLAM · TUM / 7-Scenes / EuRoC without calibration | ATE [m]: 0.114 -> 0.097 | Helped |
| use weighted-average pointmap fusion instead of median-confidence selection (ablation) | MASt3R-SLAM · TUM / 7-Scenes / EuRoC with calibration | ATE [m]: 0.039 -> 0.039 | No clear effect |
| use ray error instead of 3D point error (ablation) | uncalibrated tracking and backend optimisation · TUM / 7-Scenes / EuRoC | ATE (m): 0.155 -> 0.097 | Helped |
| add loop closure (ablation) | MASt3R-SLAM calibrated · TUM | ATE (m): 0.064 -> 0.030 | Helped |
| add loop closure (ablation) | MASt3R-SLAM calibrated · 7-Scenes | ATE (m): 0.066 -> 0.047 | Helped |
| add loop closure (ablation) | MASt3R-SLAM uncalibrated · EuRoC Vicon | Chamfer (m): 0.179 -> 0.090 | Helped |
Surprising findings
- Uncalibrated tracking with ray error gives better ATE than point error (0.097 vs 0.155 m average), and the system still beats DROID-SLAM in 7-Scenes Chamfer distance without calibration.
- On EuRoC, the calibrated system's ATE (0.041 m) trails DROID-SLAM (0.022 m), though its geometry accuracy is better.
Limitations
- Pointmaps are not refined in the full global optimisation, so geometry can be inconsistent in 3D.
- MASt3R is trained only on pinhole images, so geometry degrades with increasing lens distortion.
- Full-resolution decoder use is a bottleneck for low-latency tracking and loop-closure checks.
Experimental setup
Datasets: TUM RGB-D 7-Scenes ETH3D-SLAM EuRoC
Benchmarks: TUM RGB-D 7-Scenes ETH3D-SLAM EuRoC
Metrics: ATE RMSE AUC accuracy completion Chamfer distance FPS
Baselines: DROID-SLAM DPV-SLAM DPV-SLAM++ GO-SLAM ORB-SLAM3 DeepFactors DeepV2D NICER-SLAM Spann3R DPVO
Builds on: MASt3R DUSt3R DROID-SLAM ASMK MASt3R-SfM
Closest papers by meaning
- SLAM3R: Real-Time Dense Scene Reconstruction from Monocular RGB Videos (CVPR 2025 · similarity 0.91)
- Gaussian Splatting SLAM (CVPR 2024 · similarity 0.87)
- MUSt3R: Multi-view Network for Stereo 3D Reconstruction (CVPR 2025 · similarity 0.87)
- VGGT-SLAM: Dense RGB SLAM Optimized on the SL(4) Manifold (NeurIPS 2025 · similarity 0.86)
- DUSt3R: Geometric 3D Vision Made Easy (CVPR 2024 · similarity 0.86)
- GS-SLAM: Dense Visual SLAM with 3D Gaussian Splatting (CVPR 2024 · similarity 0.86)
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.