Research map / 3D vision / Depth and stereo
Change-Robust Online Topological Memory for Long-Term Relocalization and Semantic Navigation
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 CROSS, a topological RGB-D memory that lifts visual place recognition matches into continuous SE(3) pose hypotheses, filters them with a Gaussian-mixture belief over odometry, and commits loop closures only after persistent support.
Problem
Long-term semantic navigation needs robots to relocalize into a stored memory after appearance and scene change without corrupting the memory with false visual matches (perceptual aliasing).
Approach
A pre-commitment layer converts each retrieved keyframe into a candidate global SE(3) pose mode via RGB-D PnP, propagates competing Gaussian-mixture trajectory branches with odometry, prunes physically inconsistent branches, and promotes only persistent branches to loop closures, instead of committing place IDs or adding factors that are rejected later.
Key results
- Cross-sequence relocalization success on Rover rises to 0.397 for CROSS vs 0.158 for RTAB-Map and 0.063 for PBU; aggregate cross-sequence success 0.384.
- Topo-Bench balanced localization accuracy 0.452 for CROSS vs 0.301 for RTAB-Map (best).
- Real quadruped object-goal navigation success 70% (LC), 80% (OR), 80% (LC+OR) vs 30-60% for metric+semantic baselines, with only 10 trials per setting.
Surprising findings
- A learned relative-pose variant (VGGT) achieved zero relocalization success due to arbitrary scale across inferences.
- In the self-mapping setting, ABM achieves the highest OpenLORIS success (0.569) over CROSS (0.478).
Limitations
- Relies on keypoint matching and PnP, which can fail under extreme appearance shift, low overlap, motion blur, or noisy depth.
- Cross-sequence success remains around 40% and degrades under day-night or long-term seasonal gaps.
- Semantic evaluation uses object-query annotations/detections and a small 10-trial real-robot study, not full open-vocabulary reasoning.
Experimental setup
Datasets: OpenLORIS Corridor Rover Campus Topo-Bench
Benchmarks: OpenLORIS Corridor Rover Campus Topo-Bench real Unitree Go2 object-goal navigation
Metrics: relocalization success rate balanced localization accuracy object-goal navigation success rate Wilson confidence intervals
Baselines: ORB-SLAM3 RTAB-Map MASt3R-SLAM Greedy Matching Sequence Matching Probabilistic Belief Update ABM OpenFabmap Rat-Slam VGGT
Builds on: Gaussian-sum filter FAB-MAP RTAB-Map ORB-SLAM3 PnP
Closest papers by meaning
- MemoNav: Working Memory Model for Visual Navigation (CVPR 2024 · similarity 0.88)
- Vision-Language Reasoning for Geolocalization: A Reinforcement Learning Approach (AAAI 2026 · similarity 0.86)
- Learning with a Mole: Transferable latent spatial representations for navigation without reconstruction (ICLR 2024 · similarity 0.85)
- Kinaema: a recurrent sequence model for memory and pose in motion (NeurIPS 2025 · similarity 0.85)
- Rethinking the State Update Gate for Long-Sequence Recurrent 3D Reconstruction (NeurIPS 2026 · similarity 0.85)
- Semantic Ambiguity Modeling and Propagation for Fine-Grained Visual Cross View Geo-Localization (AAAI 2025 · 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.