Large-Scale Geo-Localization of Remote Sensing Images: A Three-Stage Framework Leveraging Maximal Clique Theory

Geo-localization of remote sensing (RS) images aims to determine the precise geographic location of a query image through image retrieval and matching techniques, with efficient and accurate localization in large-scale scenarios being a core challenge. However, existing methods face three key challenges in large-scale RS image geo-localization: 1) the disconnection between local features and global semantic modeling, leading to insufficient matching robustness; 2) single-stage retrieval strategies struggle to balance efficiency and accuracy in vast datasets; and 3) limited scene coverage in datasets, which restricts the generalization capability of algorithms. Thus, in this work, we propose a large-scale RS image geo-localization method based on maximal clique (MAC) theory, constructing a semantic-spatial-geometric three-stage optimization framework. The method sequentially incorporates semantic-guided patch-level candidate retrieval, spatio-geographic constrained MAC optimization, and MAC-guided geometric consistency for fine localization, effectively utilizing image semantics, local features, and their spatial relationships to improve localization accuracy and robustness. Additionally, we construct the RSLoc-820K dataset, which includes large-scale, continuously covered geo-referenced imagery, supporting geospatial modeling. Experimental results demonstrate that the proposed method achieves excellent performance in large-scale RS image geo-localization, particularly excelling in Top-1 accuracy. Ablation studies further validate the advantages of the MAC optimization strategy in both matching accuracy and computational efficiency, providing an effective solution for geo-localization in large-scale RS scenarios and broadening the scope of research in this field. The project will be publicly available at https://github.com/SandraPky/RSLoc-820K.