Research map / 3D vision / Gaussian splatting
VTGaussian-SLAM: RGBD SLAM for Large Scale Scenes with Splatting View-Tied 3D Gaussians
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.
Proposes VTGaussian-SLAM, an RGBD SLAM system using simplified view-tied 3D Gaussians tied to depth pixels with section-based tracking and mapping, enabling scaling to large scenes without optimizing all Gaussians.
Problem
Existing 3DGS-based RGBD SLAM methods must keep and optimize all Gaussians in limited GPU memory to maintain consistency with previous observations, which prevents scaling to extremely large scenes while keeping tracking and rendering accurate.
Approach
Tie each sphere-shaped Gaussian (color, radius, opacity only) to a depth pixel so positions follow from depth and camera pose; group consecutive frames into sections, optimize only the current section during mapping, and track head frames against a visibility-selected earlier section to limit pose error accumulation.
Key results
- On Replica tracking, average ATE RMSE of 0.28 cm, versus 0.26 for LoopSplat (which uses pretrained priors) and 0.31 for Gaussian-SLAM.
- Rendering on Replica reaches 43.34 PSNR and 0.012 LPIPS, and on TUM-RGBD 30.20 PSNR versus 25.05 for Gaussian-SLAM.
- On ScanNet++ (5 scenes), average ATE RMSE of 1.55 cm, the best among compared methods, and 32.22 PSNR rendering.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use view-tied isotropic Gaussians instead of anisotropic Gaussians with learnable locations (ablation) | VTGaussian-SLAM · Replica | PSNR: 37.62 -> 39.95 | Helped |
| Tie isotropic Gaussians to depth views instead of learning locations (ablation) | VTGaussian-SLAM · Replica | ATE RMSE: 21.45 -> 0.22 | Helped |
| Use anisotropic rather than isotropic view-tied Gaussians (ablation) | VTGaussian-SLAM · Replica | PSNR: 39.95 -> 38.46 | Hurt |
| Increase Gaussian parameters by using anisotropic view-tied Gaussians (ablation) | VTGaussian-SLAM · Replica | ATE RMSE: 0.22 -> 25.65 | Hurt |
| Set section length to 20 instead of 40 (ablation) | VTGaussian-SLAM · Replica | ATE RMSE: 0.22 -> 0.25 | Hurt |
| Set section length to 60 instead of 40 (ablation) | VTGaussian-SLAM · Replica | PSNR: 39.95 -> 39.34 | Hurt |
| Set section length to 80 instead of 40 (ablation) | VTGaussian-SLAM · Replica | PSNR: 39.95 -> 38.87 | Hurt |
| Set section length to 100 instead of 40 (ablation) | VTGaussian-SLAM · Replica | PSNR: 39.95 -> 38.56 | Hurt |
| Select the nearest section rather than the proposed overlap-selection strategy (ablation) | VTGaussian-SLAM · Replica | ATE RMSE: 0.22 -> 0.30 | Hurt |
| Select the section with the largest overlap rather than the proposed strategy (ablation) | VTGaussian-SLAM · Replica | ATE RMSE: 0.22 -> 0.34 | Hurt |
| Add the nearest section to the overlap-section selection (ablation) | VTGaussian-SLAM · Replica | PSNR: 39.95 -> 38.03 | Hurt |
| Use visibility masks instead of no visibility mask (ablation) | VTGaussian-SLAM · Replica | ATE RMSE: 6.8 -> 4.4 | Helped |
Surprising findings
- Tracking the head frame against an earlier section rather than the current section reduces pose error cumulation, even though rendering the current section gives better individual renderings.
Limitations
- Relies on depth input from RGBD sensors; tracking on ScanNet is less accurate than some priors-based methods (11.3 cm vs 7.7 cm for LoopSplat).
- ScanNet++ evaluation uses only the first 250 frames of each scene due to large motions.
- Uses somewhat more memory and runtime than some other GS-based methods (e.g., Gaussian-SLAM on Replica).
Experimental setup
Datasets: Replica TUM-RGBD ScanNet ScanNet++
Benchmarks: Replica TUM-RGBD ScanNet ScanNet++ KITTI
Metrics: ATE RMSE PSNR SSIM LPIPS F1-score D-L1
Baselines: NICE-SLAM Vox-Fusion ESLAM Point-SLAM DF-Prior Co-SLAM SplaTAM MonoGS GS-SLAM Gaussian-SLAM LoopSplat Loopy-SLAM
Builds on: 3D Gaussian Splatting SplaTAM Gaussian-SLAM
Closest papers by meaning
- GS-SLAM: Dense Visual SLAM with 3D Gaussian Splatting (CVPR 2024 · similarity 0.92)
- Gaussian Splatting SLAM (CVPR 2024 · similarity 0.91)
- SplaTAM: Splat Track & Map 3D Gaussians for Dense RGB-D SLAM (CVPR 2024 · similarity 0.90)
- DG-SLAM: Robust Dynamic Gaussian Splatting SLAM with Hybrid Pose Optimization (NeurIPS 2024 · similarity 0.90)
- VGGT-SLAM: Dense RGB SLAM Optimized on the SL(4) Manifold (NeurIPS 2025 · similarity 0.89)
- WildGS-SLAM: Monocular Gaussian Splatting SLAM in Dynamic Environments (CVPR 2025 · similarity 0.88)
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.