Research map / 3D vision / Gaussian splatting
DG-SLAM: Robust Dynamic Gaussian Splatting SLAM with Hybrid Pose Optimization
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.
DG-SLAM is a dynamic-scene RGB-D SLAM system built on 3D Gaussian splatting that combines spatio-temporal depth-warp and semantic motion masks, coarse-to-fine tracking (DROID-VO plus Gaussian-splatting photometric refinement), and adaptive Gaussian add/prune.
Problem
Camera pose estimation and static map reconstruction in dynamic scenes are not robust for Gaussian-splatting SLAM, which assumes a static environment and fails when dynamic objects cause inconsistent geometry and photometry.
Approach
Fuse a sliding-window depth-warp mask with a semantic (Oneformer) mask to filter dynamic pixels, use DROID-VO for initial pose and then refine pose via Gaussian-splatting rendering loss weighted by accumulated opacity, and add/prune Gaussians adaptively with opacity, scale and observation-count checks.
Key results
- On TUM dynamic sequences, DG-SLAM averages 2.2 cm ATE, versus 3.3 cm for DROID-VO and 5.3 cm for Rodyn-SLAM.
- On BONN dynamic sequences, DG-SLAM averages 5.5 cm ATE, versus 15.4 cm for DROID-VO and 12.3 cm for Rodyn-SLAM.
- Reconstruction on BONN gives average accuracy 8.06 cm and completion ratio 43.67% (≤5cm), better than NICE-SLAM, Co-SLAM and ESLAM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DG-SLAM instead of DROID-VO | Camera tracking · TUM RGB-D (Avg.) | ATE (cm): 3.3 -> 2.2 | Helped |
| Use DG-SLAM instead of Rodyn-SLAM | Camera tracking · BONN RGB-D Dynamic (Avg.) | ATE (cm): 12.3 -> 5.5 | Helped |
| Use DG-SLAM instead of Co-SLAM | Camera tracking · ScanNet (Avg.) | ATE (cm): 8.8 -> 8.6 | Helped |
| Use DG-SLAM instead of Co-SLAM | Static map reconstruction · BONN RGB-D Dynamic (Avg.) | Accuracy (cm): 18.06 -> 8.06 | Helped |
| Use DG-SLAM instead of ESLAM | Static map reconstruction · BONN RGB-D Dynamic (Avg.) | Completion (cm): 9.11 -> 15.46 | Hurt |
| Use DG-SLAM instead of ESLAM | Static map reconstruction · BONN RGB-D Dynamic (Avg.) | Completion Ratio (≤5 cm %): 34.93 -> 43.67 | Helped |
| Add adaptive Gaussian point addition (ablation) | DG-SLAM · BONN RGB-D Dynamic | ATE (cm): 6.63 -> 5.51 | Helped |
| Add adaptive Gaussian point pruning (ablation) | DG-SLAM · BONN RGB-D Dynamic | ATE (cm): 6.89 -> 5.51 | Helped |
| Add depth warp mask (ablation) | DG-SLAM · BONN RGB-D Dynamic | ATE (cm): 6.40 -> 5.51 | Helped |
| Add semantic segmentation mask (ablation) | DG-SLAM · BONN RGB-D Dynamic | ATE (cm): 15.27 -> 5.51 | Helped |
| Add fine Gaussian-splatting camera tracking (ablation) | DG-SLAM · BONN RGB-D Dynamic | ATE (cm): 7.36 -> 5.51 | Helped |
Surprising findings
- DG-SLAM is slower in mapping (549.3 ms per frame) than Co-SLAM (140.1 ms), although its tracking time is lowest.
Limitations
- Large-scale scene tracking and reconstruction remain a limitation; loop closure is not addressed.
- Pose accuracy depends on the segmentation precision of the semantic prior (Oneformer).
- Runs at about 2 FPS on BONN, so not real time for the whole SLAM pipeline.
Experimental setup
Datasets: TUM RGB-D BONN RGB-D Dynamic ScanNet
Benchmarks: TUM RGB-D BONN RGB-D Dynamic ScanNet
Metrics: ATE RMSE ATE STD Accuracy (cm) Completion (cm) Completion Ratio (5cm %) FPS / running time
Baselines: ORB-SLAM3 ReFusion Co-fusion MID-fusion EM-fusion iMAP NICE-SLAM Vox-Fusion Co-SLAM ESLAM Rodyn-SLAM SplaTAM
Builds on: 3D Gaussian splatting DROID-SLAM Point-SLAM ReFusion
Closest papers by meaning
- Gaussian Splatting SLAM (CVPR 2024 · similarity 0.94)
- GS-SLAM: Dense Visual SLAM with 3D Gaussian Splatting (CVPR 2024 · similarity 0.93)
- WildGS-SLAM: Monocular Gaussian Splatting SLAM in Dynamic Environments (CVPR 2025 · similarity 0.92)
- ProDyG: Progressive Dynamic Scene Reconstruction via Gaussian Splatting from Monocular Videos (NeurIPS 2025 · similarity 0.91)
- VTGaussian-SLAM: RGBD SLAM for Large Scale Scenes with Splatting View-Tied 3D Gaussians (ICML 2025 · similarity 0.90)
- SplaTAM: Splat Track & Map 3D Gaussians for Dense RGB-D SLAM (CVPR 2024 · similarity 0.89)
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.