A Transformer-Based Feature Segmentation and Region Alignment Method for UAV-View Geo-Localization

Cross-view geo-localization is a task of matching the same geographic image from different views, e.g., unmanned aerial vehicle (UAV) and satellite. The most difficult challenges are position shift uncertainty distance scale. Existing methods mainly aimed at digging for more comprehensive fine-grained information. However, it underestimates importance extracting robust feature representation impact alignment. CNN-based have achieved great success in cross-view geo-localization. However still has some limitations, can only extract part information neighborhood scale reduction operations will make lost. In particular, we introduce simple efficient transformer-based structure called Feature Segmentation Region Alignment (FSRA) to enhance model’s ability understand contextual as well distribution instances. Without using additional supervisory information, FSRA divides regions based on heat transformer’s map, then aligns multiple specific views one one. Finally, integrates each region into set representations. difference that does not divide manually, but automatically map. So instances be divided aligned when there significant shifts changes image. addition, sampling strategy proposed overcome disparity number satellite images other sources. Experiments show method superior performance achieves state-of-the-art both tasks drone view target localization navigation.