Research map / 3D vision / Depth and stereo
SplaTAM: Splat Track & Map 3D Gaussians for Dense RGB-D 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.
Introduces SplaTAM, a dense RGB-D SLAM system that represents the scene as isotropic 3D Gaussians and jointly tracks camera pose and maps via silhouette-guided differentiable splatting.
Problem
Dense RGB-D SLAM with explicit, high-fidelity, and novel-view-capable maps; prior implicit neural radiance-field SLAM is slow and hard to edit, and traditional representations lack rendering and photometric optimization.
Approach
Uses 3D Gaussian Splatting as the sole SLAM map, with simplified isotropic view-independent Gaussians, a rendered silhouette to mask tracking to well-mapped regions, silhouette-based densification of unmapped areas, and dense photometric plus depth losses for tracking and mapping.
Key results
- Achieves 0.36 cm average ATE RMSE on Replica versus 0.52 cm for Point-SLAM, and 5.48 cm on TUM-RGBD versus 8.92 cm for Point-SLAM.
- Tracks ScanNet++ sequences with 1.2 cm average ATE RMSE, while Point-SLAM and ORB-SLAM3 fail (343.8 cm and 158.2 cm).
- Novel-view synthesis on ScanNet++ reaches 24.41 dB PSNR, where Point-SLAM reaches 11.91 dB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SplaTAM instead of ORB-SLAM3 | ScanNet++ camera tracking · ScanNet++ | ATE RMSE [cm]: 158.2 -> 1.2 | Helped |
| use SplaTAM instead of Point-SLAM | Replica camera tracking · Replica | ATE RMSE [cm]: 0.52 -> 0.36 | Helped |
| use SplaTAM instead of ORB-SLAM2 | TUM-RGBD camera tracking · TUM-RGBD | ATE RMSE [cm]: 1.98 -> 5.48 | Hurt |
| use SplaTAM instead of NICE-SLAM | Orig-ScanNet camera tracking · Orig-ScanNet | ATE RMSE [cm]: 10.70 -> 11.88 | Hurt |
| use SplaTAM instead of Point-SLAM | Replica training-view rendering · Replica | PSNR: 35.17 -> 34.11 | Hurt |
| remove forward velocity propagation (ablation) | SplaTAM camera tracking · Replica Room 0 | ATE [cm]: 0.27 -> 2.95 | Hurt |
| remove silhouette mask (ablation) | SplaTAM camera tracking · Replica Room 0 | ATE [cm]: 0.27 -> 115.80 | Hurt |
| use silhouette threshold 0.5 instead of 0.99 (ablation) | SplaTAM camera tracking · Replica Room 0 | ATE [cm]: 0.27 -> 1.30 | Hurt |
| use silhouette threshold 0.5 instead of 0.99 (ablation) | SplaTAM mapping · Replica Room 0 | PSNR [dB]: 32.81 -> 31.36 | Hurt |
| remove RGB loss while retaining depth loss (ablation) | SplaTAM tracking and mapping · Replica Room 0 | ATE [cm]: null | Hurt |
| remove depth loss while retaining RGB loss (ablation) | SplaTAM tracking and mapping · Replica Room 0 | ATE [cm]: null | Hurt |
| use only RGB loss instead of RGB and depth losses (ablation) | SplaTAM mapping · Replica Room 0 | PSNR: 1.5 PSNR lower | Hurt |
Surprising findings
- Using only a depth loss causes tracking to fail completely, while only an RGB loss gives more than 5x the error of using both.
- Point-SLAM's rendering results on Replica are comparable to SplaTAM's despite Point-SLAM using ground-truth depth to sample rendering points.
Limitations
- Sensitive to motion blur, large depth noise, and aggressive rotation.
- Requires known camera intrinsics and dense depth input.
- Scaling to large scenes is left to future work (e.g., OpenVDB).
Experimental setup
Datasets: ScanNet++ Replica TUM-RGBD ScanNet
Benchmarks: ScanNet++ Replica TUM-RGBD ScanNet
Metrics: ATE RMSE PSNR SSIM LPIPS Depth L1
Baselines: Point-SLAM NICE-SLAM Vox-Fusion ESLAM ORB-SLAM3 DROID-SLAM Kintinuous ElasticFusion ORB-SLAM2
Builds on: 3D Gaussian Splatting Point-SLAM NICE-SLAM
Closest papers by meaning
- Gaussian Splatting SLAM (CVPR 2024 · similarity 0.92)
- GS-SLAM: Dense Visual SLAM with 3D Gaussian Splatting (CVPR 2024 · similarity 0.92)
- VTGaussian-SLAM: RGBD SLAM for Large Scale Scenes with Splatting View-Tied 3D Gaussians (ICML 2025 · similarity 0.90)
- DG-SLAM: Robust Dynamic Gaussian Splatting SLAM with Hybrid Pose Optimization (NeurIPS 2024 · similarity 0.89)
- IncEventGS: Pose-Free Gaussian Splatting from a Single Event Camera (CVPR 2025 · similarity 0.87)
- SplatFormer: Point Transformer for Robust 3D Gaussian Splatting (ICLR 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.