Research map / 3D vision / Gaussian splatting
Gaussian Splatting SLAM
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.
Presents Gaussian Splatting SLAM, the first monocular and RGB-D SLAM system using 3D Gaussians as the sole scene representation, running live at about 3 fps.
Problem
Online visual SLAM needs a unified dense map representation that supports accurate camera tracking, efficient mapping, and high-fidelity rendering, especially in the hard monocular setting without loop closure.
Approach
Use 3D Gaussian Splatting as the only SLAM representation, with an analytic Lie-group Jacobian of camera pose w.r.t. the Gaussians for direct photometric tracking, plus isotropic scale regularisation, covisibility-based keyframe management, and Gaussian insertion/pruning for monocular geometric verification.
Key results
- Monocular ATE RMSE on TUM RGB-D averages 3.96 cm, beating DSO (11.0 cm) and DROID-VO (7.73 cm) without deep priors or loop closure.
- On Replica RGB-D novel view synthesis, PSNR 38.94 dB vs 35.17 for Point-SLAM, with rendering at 769 FPS.
- RGB-D tracking on TUM averages 1.47 cm ATE RMSE, the best among the compared neural-implicit SLAM methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add isotropic regularisation (ablation) | Gaussian Splatting SLAM · TUM RGB-D, monocular | ATE RMSE (cm): 4.83 -> 3.96 | Helped |
| add keyframe selection (ablation) | Gaussian Splatting SLAM · TUM RGB-D, monocular | ATE RMSE (cm): 8.73 -> 3.96 | Helped |
| add geometric residual (ablation) | Gaussian Splatting SLAM · TUM RGB-D, RGB-D | ATE RMSE (cm): 2.66 -> 1.47 | Helped |
| add keyframe selection (ablation) | Gaussian Splatting SLAM · TUM RGB-D, RGB-D | ATE RMSE (cm): 1.90 -> 1.47 | Helped |
| use depth for map training (ablation) | Gaussian Splatting SLAM · Camera convergence sequences | Successful convergence ratio: 0.79 -> 0.82 | Helped |
| use Gaussian Splatting SLAM instead of Point-SLAM | Replica RGB-D rendering · Replica RGB-D | PSNR[db]: 35.17 -> 38.94 | Helped |
| use Gaussian Splatting SLAM instead of Point-SLAM | Replica RGB-D rendering · Replica RGB-D | SSIM: 0.975 -> 0.968 | Hurt |
| use Gaussian Splatting SLAM instead of Point-SLAM | Replica RGB-D rendering · Replica RGB-D | LPIPS: 0.124 -> 0.070 | Helped |
| use Gaussian Splatting SLAM instead of Point-SLAM | Replica RGB-D rendering · Replica RGB-D | Rendering FPS: 1.33 -> 769 | Helped |
Surprising findings
- Gaussian maps show a much larger camera convergence basin than hash-grid or MLP SDF representations in the convergence funnel analysis.
Limitations
- No loop closure, so large-scale scenes are not handled.
- Geometry such as surface normals is not explicitly extracted because Gaussians do not represent surfaces explicitly.
- Monocular tracking lacks deep priors and depends on photometric signals; Replica RGB-D single-process results trade off against multi-process speed.
Experimental setup
Datasets: TUM RGB-D Replica
Benchmarks: TUM RGB-D Replica
Metrics: ATE RMSE PSNR SSIM LPIPS rendering FPS convergence success rate
Baselines: DSO DepthCov-VO DROID-SLAM ORB-SLAM2 iMAP NICE-SLAM Point-SLAM Co-SLAM ESLAM Vox-Fusion
Builds on: 3D Gaussian Splatting EWA splatting DSO Lie group pose optimisation
Closest papers by meaning
- GS-SLAM: Dense Visual SLAM with 3D Gaussian Splatting (CVPR 2024 · similarity 0.94)
- DG-SLAM: Robust Dynamic Gaussian Splatting SLAM with Hybrid Pose Optimization (NeurIPS 2024 · similarity 0.94)
- SplaTAM: Splat Track & Map 3D Gaussians for Dense RGB-D SLAM (CVPR 2024 · similarity 0.92)
- VTGaussian-SLAM: RGBD SLAM for Large Scale Scenes with Splatting View-Tied 3D Gaussians (ICML 2025 · similarity 0.91)
- KeyGS: A Keyframe-Centric Gaussian Splatting Method for Monocular Image Sequences (AAAI 2025 · similarity 0.91)
- IncEventGS: Pose-Free Gaussian Splatting from a Single Event Camera (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.