MAG-VLAQ matches ground observations (camera and LiDAR) against aerial or satellite images for place recognition. Foundation models give dense visual tokens from both views and geometric tokens from ground LiDAR, and a neural-ODE fusion of RGB and LiDAR adapts the query centres of a VLAQ aggregator to each scene. On KITTI360-AG it reaches 61.1 Recall@1 in the satellite setting, against 34.5 for the next best method.
Relevance
Overhead imagery is a reference a mixed team can fall back on when GPS is weak or spoofed. Cross-view multi-robot multimodal perception, related to map alignment in collaborative SLAM.
Links
- arXiv
- Subfield: Field robotics, Mapping and localisation, Perception