GeoUniPR matches LiDAR and camera data with geometry-consistent place recognition

GeoUniPR, developed by Wonbong Kim and seven colleagues, tackles cross-modal place recognition (CMPR) — matching the same location across vision and LiDAR sensors. Existing methods rely on complex alignment modules, multi-stage training, or full fine-tuning of pretrained backbones. GeoUniPR simplifies this by revisiting the problem from the angle of geometric consistency. The framework first projects LiDAR point clouds into the camera's perspective to create depth-image views (DIV), establishing direct RGB-LiDAR correspondence. These views are further enriched with native LiDAR cues — intensity and surface-normal information — producing a multi-channel geometric representation that strengthens structural consistency across modalities.

With this representation, GeoUniPR learns a unified embedding space using two ViT-based encoders with identical architectures, trained through parameter-efficient adaptation — no auxiliary alignment modules, no multi-stage pipelines, and no full backbone fine-tuning. The authors also introduce Spatially-Consistent InfoNCE (SC-InfoNCE), a contrastive objective tailored for CMPR that suppresses false negatives caused by spatial continuity in continuous scenes. Experiments on the KITTI and KITTI-360 datasets show GeoUniPR achieves state-of-the-art performance in both same-modal and cross-modal place recognition, with strong cross-dataset generalization. The approach's simplicity and efficiency make it a practical candidate for real-time autonomous driving and robotics localization systems.