Research map / 3D vision / Depth and stereo
SLAM3R: Real-Time Dense Scene Reconstruction from Monocular RGB Videos
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.
SLAM3R performs real-time dense 3D reconstruction from monocular RGB video by feed-forward regression of local pointmaps in sliding windows and incremental registration into a global frame, without explicit camera solving.
Problem
Dense 3D scene reconstruction from monocular RGB video that is accurate, complete, and real-time, avoiding the slow optimization and camera-pose solving of existing dense SLAM and multi-view pipelines.
Approach
Two feed-forward networks, an Image-to-Points (I2P) network that regresses keyframe pointmaps for sliding-window clips and a Local-to-World (L2W) network that registers them into a global frame using retrieved scene frames from a reservoir, extend DUSt3R to multi-view with no explicit camera parameter estimation or global optimization.
Key results
- On 7 Scenes, SLAM3R achieves average accuracy/completeness of 2.13/2.34 cm at ~25 FPS, beating DUSt3R, MASt3R and Spann3R (3.24/3.90/2.41 cm for completeness/accuracy comparisons as reported).
- On Replica, SLAM3R averages 3.57/2.62 cm accuracy/completeness at ~24 FPS, surpassing all baselines running above 1 FPS, including Spann3R (10.32/13.33 cm).
- Pose ATE-RMSE on 7 Scenes is 8.41 cm for SLAM3R versus 5.66 cm for DROID-SLAM, indicating pose accuracy lags while reconstruction is strong.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SLAM3R instead of Spann3R | dense reconstruction · 7 Scenes | accuracy (cm): 3.42 -> 2.13 | Helped |
| use SLAM3R instead of Spann3R | dense reconstruction · 7 Scenes | completeness (cm): 2.41 -> 2.34 | Helped |
| use SLAM3R instead of DROID-SLAM | dense reconstruction · Replica | accuracy (cm): 5.50 -> 3.57 | Helped |
| use SLAM3R instead of DROID-SLAM | dense reconstruction · Replica | completeness (cm): 12.29 -> 2.62 | Helped |
| increase the I2P window from 2 to 11 frames (ablation) | I2P keyframe reconstruction · Replica | accuracy: 3.39 -> 2.38 | Helped |
| increase the I2P window from 2 to 11 frames (ablation) | I2P keyframe reconstruction · Replica | completeness: 3.04 -> 2.03 | Helped |
| increase the I2P window from 11 to 51 frames (ablation) | I2P keyframe reconstruction · Replica | accuracy: 2.38 -> 2.23 | Helped |
| increase the I2P window from 11 to 51 frames (ablation) | I2P keyframe reconstruction · Replica | FPS: 40.11 -> 11.97 | Hurt |
| filter pointmaps with a confidence threshold of 3 (ablation) | SLAM3R · 7 Scenes | accuracy (cm): 2.40 -> 2.13 | Helped |
| filter pointmaps with a confidence threshold of 3 (ablation) | SLAM3R · 7 Scenes | completeness (cm): 2.24 -> 2.34 | Hurt |
| add the L2W registration network instead of global optimization (ablation) | I2P alignment · Replica | accuracy: 4.87 -> 3.62 | Helped |
| add retrieval of correlated historical scene frames (ablation) | I2P+L2W · Replica | accuracy: 6.19 -> 3.62 | Helped |
Surprising findings
- Camera pose error and reconstruction quality are not fully positively correlated, suggesting end-to-end reconstruction can work without precise poses.
- Pose accuracy (ATE) of SLAM3R is worse than DROID-SLAM on Replica (6.61 cm vs 0.33 cm), even though reconstruction is better.
Limitations
- Eliminating camera parameter prediction prevents global bundle adjustment.
- Poses derived from the scene point cloud are less accurate than dedicated SLAM camera localization systems.
Experimental setup
Datasets: ScanNet++ Aria Synthetic Environments CO3D-v2
Benchmarks: 7 Scenes Replica
Metrics: accuracy (cm) completeness (cm) FPS ATE-RMSE (cm)
Baselines: DUSt3R MASt3R Spann3R NICER-SLAM DROID-SLAM DIM-SLAM GO-SLAM
Builds on: DUSt3R Spann3R MASt3R
Closest papers by meaning
- MASt3R-SLAM: Real-Time Dense SLAM with 3D Reconstruction Priors (CVPR 2025 · similarity 0.91)
- VGGT-SLAM: Dense RGB SLAM Optimized on the SL(4) Manifold (NeurIPS 2025 · similarity 0.88)
- Point3R: Streaming 3D Reconstruction with Explicit Spatial Pointer Memory (NeurIPS 2025 · similarity 0.88)
- Benchmarking Implicit Neural Representation and Geometric Rendering in Real-Time RGB-D SLAM (CVPR 2024 · similarity 0.88)
- MUSt3R: Multi-view Network for Stereo 3D Reconstruction (CVPR 2025 · similarity 0.87)
- ProDyG: Progressive Dynamic Scene Reconstruction via Gaussian Splatting from Monocular Videos (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.