Research map / 3D vision / Depth and stereo
Loopy-SLAM: Dense Neural SLAM with Loop Closures
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.
Loopy-SLAM adds submap-based loop closure with pose graph optimization to a point-based neural RGBD SLAM system (Point-SLAM), enabling drift-corrected dense mapping and tracking without storing input frames.
Problem
Dense neural RGBD SLAM suffers from error accumulation during camera tracking, producing distorted maps, and existing loop-closure approaches with hash grids require gradient-based re-optimization and storing all mapped frames.
Approach
Anchor neural features on point-cloud submaps that are created dynamically from camera motion, detect loops with BoW place recognition over global keyframes, and correct poses and submaps by rigid transforms from robust pose graph optimization over dense surface registration, so the point-based map can be shifted without retraining.
Key results
- On Replica tracking, Loopy-SLAM averages 0.29 cm ATE RMSE vs 0.35 cm for GO-SLAM and 0.52 cm for Point-SLAM.
- On ScanNet, average ATE over 9 scenes is 7.7 cm vs 7.0 cm for GO-SLAM; the paper reports it beats all other dense neural methods on average.
- Replica reconstruction average depth L1 is 0.35 cm vs 0.44 cm for Point-SLAM, with F1 90.8% vs 89.8%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add online loop closure and global pose-graph optimization to the point-based SLAM system | Loopy-SLAM · Replica | ATE RMSE (cm): 0.35 -> 0.29 | Helped |
| Add online loop closure and global pose-graph optimization to the point-based SLAM system | Loopy-SLAM · Replica | Depth L1 (cm): 0.44 -> 0.35 | Helped |
| Add online loop closure and global pose-graph optimization to the point-based SLAM system | Loopy-SLAM · Replica | F1 (%): 89.8 -> 90.8 | Helped |
| Add online loop closure and global pose-graph optimization to the point-based SLAM system | Loopy-SLAM · TUM-RGBD | ATE RMSE (cm): 7.89 -> 3.85 | Helped |
| Add online loop closure and global pose-graph optimization to the point-based SLAM system | Loopy-SLAM · TUM-RGBD fr1/room | ATE RMSE (cm): 30.92 -> 7.03 | Helped |
| Add online loop closure and global pose-graph optimization to the point-based SLAM system | Loopy-SLAM · ScanNet Avg.-6 | ATE RMSE (cm): 6.9 -> 7.7 | Hurt |
| Add online loop closure and global pose-graph optimization to the point-based SLAM system | Loopy-SLAM · ScanNet Avg.-9 | ATE RMSE (cm): 7.0 -> 7.7 | Hurt |
| Add online loop closure and global pose-graph optimization to the point-based SLAM system | Loopy-SLAM · Replica | PSNR (dB): 35.17 -> 35.47 | Helped |
| Add online loop closure and global pose-graph optimization to the point-based SLAM system | Loopy-SLAM · Replica | SSIM: 0.975 -> 0.981 | Helped |
| Add online loop closure and global pose-graph optimization to the point-based SLAM system | Loopy-SLAM · Replica | LPIPS: 0.124 -> 0.109 | Helped |
Surprising findings
- On ScanNet, Loopy-SLAM's average ATE (7.7 cm) is slightly worse than GO-SLAM's (7.0 cm) despite beating other dense neural methods.
Limitations
- Tracker is frame-to-model only, prone to drift; not robust enough for noisy real-world data per authors.
- Implementation in PyTorch and Open3D is not optimized for real-time operation; registrations are slow.
- No relocalization implemented.
Experimental setup
Datasets: Replica TUM-RGBD ScanNet
Benchmarks: Replica TUM-RGBD ScanNet
Metrics: ATE RMSE depth L1 F-score PSNR SSIM LPIPS
Baselines: Point-SLAM ESLAM GO-SLAM NICE-SLAM Vox-Fusion MIPS-Fusion Co-SLAM ORB-SLAM2
Builds on: Point-SLAM Choi et al. robust pose graph optimization TSDF Fusion bag of visual words
Closest papers by meaning
- Benchmarking Implicit Neural Representation and Geometric Rendering in Real-Time RGB-D SLAM (CVPR 2024 · similarity 0.89)
- IBD-SLAM: Learning Image-Based Depth Fusion for Generalizable SLAM (CVPR 2024 · similarity 0.88)
- GS-SLAM: Dense Visual SLAM with 3D Gaussian Splatting (CVPR 2024 · similarity 0.87)
- VTGaussian-SLAM: RGBD SLAM for Large Scale Scenes with Splatting View-Tied 3D Gaussians (ICML 2025 · similarity 0.87)
- WildGS-SLAM: Monocular Gaussian Splatting SLAM in Dynamic Environments (CVPR 2025 · 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.