Focused Region Exclusion Learning for Cross-View Geolocalization

Cross-view geolocalization aims to establish matching relationships between images of the same scene captured from different viewpoints. However, significant viewpoint discrepancies can render some visual information ineffective across views, and complex scene structures lead to dispersed retrieval-relevant features, which together pose substantial challenges to geolocalization. To address these issues, existing methods often rely on alignment-based strategies to reduce viewpoint discrepancies and suppress irrelevant information during retrieval. Nevertheless, such methods typically exhibit limited generalizability across diverse scenarios and often fail to effectively handle complex visual scenes. To fully leverage effective visual information, we propose a novel vision-centric method called focused region exclusion learning. The method employs primary and auxiliary branch types, in which the auxiliary branch uses attended regions of the primary branch as guidance and actively explores complementary cues beyond those regions, thereby uncovering discriminative visual information overlooked by the primary branch and enhancing the diversity and robustness of the learned features. To achieve dynamic modulation of visual attention across branches, we propose a heatmap-guided mask filtering module. In addition, a cross-guided learning loss is incorporated to facilitate effective feature fusion between the different branches. The results of experiments on four benchmark datasets CVUSA, CVACT, VIGOR, and University-1652 demonstrate that the proposed method exhibits strong dataset applicability and consistently outperforms state-of-The-Art methods across these benchmarks.