Research map / 3D vision / Gaussian splatting
GS-SLAM: Dense Visual SLAM with 3D Gaussian Splatting
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.
Introduces GS-SLAM, the first dense RGB-D SLAM system built on 3D Gaussian splatting, with adaptive Gaussian expansion/deletion for mapping and coarse-to-fine tracking, running at about 8 FPS while rendering at about 386 FPS on Replica.
Problem
Dense RGB-D SLAM with neural implicit representations trades off efficiency and accuracy: ray-based volume rendering is slow and only a few pixels are rendered per optimization step, yielding less detailed maps. 3D Gaussian splatting had not been applied to camera tracking or real-time mapping.
Approach
Represent the scene as explicit 3D Gaussians rendered by differentiable splatting, add an adaptive expansion/deletion strategy that inserts Gaussians in unreliable (new) regions and removes floaters by depth, and use a coarse-to-fine tracking scheme that first optimizes pose on sparse pixels and then on full-resolution renders from reliable Gaussians selected by depth consistency.
Key results
- On Replica, tracking ATE averages 0.50 cm, outperforming the second-best Point-SLAM (0.54 cm) while running at 8.34 FPS versus 0.42 FPS for Point-SLAM.
- Rendering on Replica reaches 34.27 dB PSNR and 386 FPS average, about 100x faster than Vox-Fusion and 1.52 dB PSNR above the second-best CoSLAM.
- Mapping achieves the best average Depth L1 (1.16 cm) and Precision (74.0%) on Replica; on TUM-RGBD average ATE is 3.7 cm, comparable to SOTA neural SLAM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use GS-SLAM instead of Point-SLAM | Replica localization · Replica | ATE RMSE (cm): 0.54 -> 0.50 | Helped |
| use GS-SLAM instead of CoSLAM | Replica rendering · Replica | PSNR (dB): +1.52 dB | Helped |
| use GS-SLAM instead of CoSLAM | Replica rendering · Replica | SSIM: 0.939 -> 0.975 | Helped |
| use GS-SLAM instead of NICE-SLAM | Replica rendering · Replica | LPIPS: 0.233 -> 0.082 | Helped |
| use GS-SLAM instead of Vox-Fusion | Replica rendering · Replica | rendering FPS: 386 FPS; 100⇥ faster | Helped |
| add adaptive Gaussian expansion, including adding and deleting Gaussians (ablation) | GS-SLAM mapping · Replica #Room0 | ATE: 0.58 -> 0.48 | Helped |
| add adaptive Gaussian expansion, including adding and deleting Gaussians (ablation) | GS-SLAM mapping · Replica #Room0 | Recall: 49.32 -> 61.29 | Helped |
| remove adaptive Gaussian adding (ablation) | GS-SLAM mapping · Replica #Room0 | mapping performance | Hurt |
| remove adaptive Gaussian deletion (ablation) | GS-SLAM mapping · Replica #Room0 | ATE: 0.48 -> 0.58 | Hurt |
| use coarse-to-fine tracking instead of fine tracking only (ablation) | GS-SLAM tracking · Replica #Room0 | ATE: 0.49 -> 0.48 | Helped |
| use coarse-to-fine tracking instead of coarse tracking only (ablation) | GS-SLAM tracking · Replica #Room0 | ATE: 0.91 -> 0.48 | Helped |
| use coarse-to-fine tracking instead of fine tracking only (ablation) | GS-SLAM tracking · Replica #Room0 | Recall: 59.18 -> 61.29 | Helped |
Surprising findings
- Without the adding step, the system completely crashes because 3DGS density control cannot handle real-time mapping without an accurate initial point cloud.
- The 3D Gaussian representation uses about 198 MB memory, roughly 4x NICE-SLAM, mainly due to spherical harmonic coefficients.
Limitations
- Relies on high-quality depth data, which may limit performance in some conditions.
- High memory requirements for large scenes due to spherical harmonic coefficients.
- Still a gap to traditional SLAM systems on TUM-RGBD tracking accuracy.
Experimental setup
Datasets: Replica TUM-RGBD
Benchmarks: Replica TUM-RGBD
Metrics: ATE RMSE Depth L1 Precision Recall F1 PSNR SSIM LPIPS FPS
Baselines: NICE-SLAM Vox-Fusion CoSLAM ESLAM Point-SLAM iMAP
Builds on: 3D Gaussian Splatting NICE-SLAM iMAP
Closest papers by meaning
- Gaussian Splatting SLAM (CVPR 2024 · similarity 0.94)
- DG-SLAM: Robust Dynamic Gaussian Splatting SLAM with Hybrid Pose Optimization (NeurIPS 2024 · similarity 0.93)
- VTGaussian-SLAM: RGBD SLAM for Large Scale Scenes with Splatting View-Tied 3D Gaussians (ICML 2025 · similarity 0.92)
- SplaTAM: Splat Track & Map 3D Gaussians for Dense RGB-D SLAM (CVPR 2024 · similarity 0.92)
- WildGS-SLAM: Monocular Gaussian Splatting SLAM in Dynamic Environments (CVPR 2025 · similarity 0.90)
- POp-GS: Next Best View in 3D-Gaussian Splatting with P-Optimality (CVPR 2025 · similarity 0.90)
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.