GeoISF boosts LiDAR-to-satellite localization 13x with semantic forest

Cross-view geo-localization—matching a ground LiDAR scan to an aerial satellite image—remains a key challenge for autonomous vehicles and robotics, especially at city scale. Existing methods struggle with the deep modality gap between sparse 3D point clouds and dense 2D satellite imagery, and often fail to capture long-term temporal semantics. In a new paper, Di Hu, Xia Yuan, and Chunxia Zhao present GeoISF (Instance Semantic Forest), a pipeline that leverages a hierarchical semantic forest constructed from WordNet. By integrating semantic trees from multiple frames, GeoISF enriches temporal representation and discriminative power, enabling more robust alignment across modalities.

Extensive experiments on the KITTI dataset show a dramatic improvement: GeoISF achieves 13.22x higher recall at Rank-1 (R@10) compared to the previous state-of-the-art LiDAR-to-image method. The approach effectively bridges the semantic gap by using environmental semantics as a shared medium, and handles large-scale scenarios without the computational explosion typical of brute-force retrieval. The authors plan to release the code as open source, which could accelerate development of practical localization systems for autonomous driving, drone navigation, and urban mapping where GPS is unreliable or denied.