Research map / 3D vision / Depth and stereo
PLGSLAM: Progressive Neural Scene Represenation with Local to Global Bundle Adjustment
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 PLGSLAM, a neural RGB-D SLAM system that progressively allocates local tri-plane+MLP scene representations and uses local-to-global bundle adjustment with a global keyframe database for accurate large-scale reconstruction and tracking.
Problem
Neural implicit SLAM methods use a single fixed-capacity global radiance field and end-to-end pose networks, which yield poor reconstruction and drifting camera tracking in large indoor scenes and long sequences.
Approach
Dynamically allocate new local scene representations (tri-planes for high-frequency detail plus one-blob-encoded MLPs for smooth low-frequency completion) as the camera leaves the current bounds, and refine poses with a local-to-global bundle adjustment that uses neural warping and SIFT reprojection losses over a global keyframe database.
Key results
- On the Apartment dataset, PLGSLAM reaches ATE RMSE 6.228 cm versus 6.891 cm for Co-SLAM and 7.283 cm for ESLAM.
- On Replica, PLGSLAM achieves ATE RMSE 0.635 cm and completion ratio 97.877%, outperforming ESLAM (0.707 cm, 96.427%).
- On ScanNet, PLGSLAM reports ATE RMSE 6.77 cm versus 8.75 cm for Co-SLAM, with average frame processing time 0.37 s versus 0.38 s for Co-SLAM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use joint tri-plane and MLP encoding (ablation) | PLGSLAM · Apartment | Accuracy [cm]: 13.314 -> 11.086 | Helped |
| Use progressive scene representation instead of a single network (ablation) | PLGSLAM · Apartment | Accuracy [cm]: 12.754 -> 11.086 | Helped |
| Use local-to-global bundle adjustment (ablation) | PLGSLAM · Apartment | Accuracy [cm]: 12.435 -> 11.086 | Helped |
Limitations
- Global BA and progressive allocation add complexity; gains on small Replica rooms are modest.
- Culling strategy for evaluation is acknowledged as removing many meshes, leaving evaluation sensitive to protocol.
Experimental setup
Datasets: Replica ScanNet Apartment (NICE-SLAM)
Benchmarks: Replica ScanNet Apartment
Metrics: Depth L1 Accuracy Completion Completion ratio ATE mean ATE RMSE frame processing time
Baselines: iMAP NICE-SLAM Vox-Fusion ESLAM Co-SLAM
Builds on: NICE-SLAM ESLAM Co-SLAM NeRF Mega-NeRF
Closest papers by meaning
- WildGS-SLAM: Monocular Gaussian Splatting SLAM in Dynamic Environments (CVPR 2025 · similarity 0.85)
- ProDyG: Progressive Dynamic Scene Reconstruction via Gaussian Splatting from Monocular Videos (NeurIPS 2025 · similarity 0.85)
- SLAM3R: Real-Time Dense Scene Reconstruction from Monocular RGB Videos (CVPR 2025 · similarity 0.85)
- Continuous Pose for Monocular Cameras in Neural Implicit Representation (CVPR 2024 · similarity 0.85)
- Loopy-SLAM: Dense Neural SLAM with Loop Closures (CVPR 2024 · similarity 0.85)
- IBD-SLAM: Learning Image-Based Depth Fusion for Generalizable SLAM (CVPR 2024 · similarity 0.85)
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.