Research map / 3D vision / Gaussian splatting
WildGS-SLAM: Monocular Gaussian Splatting SLAM in Dynamic Environments
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 WildGS-SLAM, a monocular Gaussian-splatting SLAM system that predicts per-pixel uncertainty from DINOv2 features with a shallow MLP to down-weight dynamic distractors in dense bundle adjustment and Gaussian map optimization.
Problem
Monocular RGB SLAM in dynamic, in-the-wild environments where moving objects break tracking and corrupt dense reconstruction, without relying on depth sensors or predefined semantic classes.
Approach
Train an online uncertainty MLP on 3D-aware DINOv2 features (with NeRF On-the-go style losses plus a depth loss from Metric3D v2) to weight the DBA objective for tracking and the rendering loss for mapping, while detaching gradients between the uncertainty predictor and the Gaussian map.
Key results
- On the Wild-SLAM MoCap dataset, WildGS-SLAM reaches 0.46 cm average ATE RMSE versus 2.40 cm for MegaSaM and 8.71 cm for Splat-SLAM.
- On the Bonn RGB-D Dynamic dataset, WildGS-SLAM achieves 2.31 cm average ATE RMSE, the best among monocular and RGB-D methods reported.
- Novel view synthesis on Wild-SLAM MoCap reaches 20.59 average PSNR and 0.783 SSIM, above Splat-SLAM's 17.23 PSNR and 0.699 SSIM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use uncertainty-aware tracking and mapping | WildGS-SLAM · Wild-SLAM MoCap | ATE RMSE (cm): 2.40 -> 0.46 | Helped |
| use uncertainty-aware tracking and mapping | WildGS-SLAM · Bonn RGB-D Dynamic Dataset | ATE RMSE (cm): 2.91 -> 2.31 | Helped |
| use uncertainty-aware tracking and mapping | WildGS-SLAM · TUM RGB-D | ATE RMSE (cm): 1.63 -> 1.63 | No clear effect |
| use uncertainty-aware mapping | WildGS-SLAM · Wild-SLAM MoCap | PSNR: 17.23 -> 20.59 | Helped |
| add an uncertainty mask (ablation) | WildGS-SLAM · Wild-SLAM MoCap | ATE RMSE (cm): 3.89 -> 0.46 | Helped |
| add the L1 depth loss to uncertainty training (ablation) | WildGS-SLAM · Wild-SLAM MoCap | ATE RMSE (cm): 0.50 -> 0.46 | Helped |
| use the learned uncertainty mask instead of YOLOv8 + SAM masks (ablation) | WildGS-SLAM · Wild-SLAM MoCap | ATE RMSE (cm): 3.06 -> 0.46 | Helped |
| add disparity regularization (ablation) | WildGS-SLAM · Wild-SLAM MoCap | ATE RMSE (cm): 10.97 -> 0.46 | Helped |
| add disparity regularization (ablation) | WildGS-SLAM · Bonn RGB-D Dynamic Dataset | ATE RMSE (cm) | Helped |
Surprising findings
- Removing the L1 depth loss from the uncertainty objective yields slightly lower Bonn error on some sequences but a much larger Wild-SLAM error gap is not the case; the depth loss mainly helps distractor separation.
- The naive MonST3R sliding-window adaptation leads to large tracking errors on long sequences despite strong short-sequence performance.
Limitations
- Uncertainty predictor is trained online, so it struggles to recognize distractors when few views cover a region.
- Relies on monocular metric depth from Metric3D v2 and DINOv2 features, with blurry uncertainty maps due to 1/14 feature resolution.
Experimental setup
Datasets: Bonn RGB-D Dynamic Dataset TUM RGB-D Dataset Wild-SLAM MoCap Dataset Wild-SLAM iPhone Dataset
Benchmarks: Wild-SLAM MoCap Dataset Bonn RGB-D Dynamic Dataset TUM RGB-D Dataset Wild-SLAM iPhone Dataset
Metrics: ATE RMSE PSNR SSIM LPIPS
Baselines: DSO ORB-SLAM2 DROID-SLAM ReFusion DynaSLAM NICE-SLAM MonoGS Splat-SLAM DG-SLAM RoDyn-SLAM DDN-SLAM DynaMoN
Builds on: DROID-SLAM MonoGS Splat-SLAM NeRF On-the-go WildGaussians
Closest papers by meaning
- DG-SLAM: Robust Dynamic Gaussian Splatting SLAM with Hybrid Pose Optimization (NeurIPS 2024 · similarity 0.92)
- Gaussian Splatting SLAM (CVPR 2024 · similarity 0.90)
- GS-SLAM: Dense Visual SLAM with 3D Gaussian Splatting (CVPR 2024 · similarity 0.90)
- ProDyG: Progressive Dynamic Scene Reconstruction via Gaussian Splatting from Monocular Videos (NeurIPS 2025 · similarity 0.90)
- VTGaussian-SLAM: RGBD SLAM for Large Scale Scenes with Splatting View-Tied 3D Gaussians (ICML 2025 · similarity 0.88)
- VGGT-SLAM: Dense RGB SLAM Optimized on the SL(4) Manifold (NeurIPS 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.