Do Keypoints Contain Crucial Information? Mining Keypoint Information to Enhance Cross-View Geo-Localization

Due to drastic view changes and different capturing times between images, extracting discriminative image-level features for cross-view geo-localization is challenging. Although recent works have achieved outstanding progress on cross-view geo-localization, the fine-grained information in images has not been fully explored in extracting image-level features. Inspired by the process of the human visual system to distinguish similar targets and the process of keypoint detection and description, we propose a framework called UDPA-Net, which guides the model to mine more favorable information for cross-view geolocalization by detecting keypoints. Specifically, we design a Unit Dot Product Attention Module (UDPAM) to discover remarkable keypoints automatically and guide the model to pay more attention to the salient regions. UDPA-Net introduces few parameters but yields significant performance gains and can be easily integrated into different networks. Our code is available at https://gitee.com/KerasLyc/UDPA-Net.