Abstract: To solve the problem that cross-view images have great viewpoint difference and thus resulting in low geolocalization accuracy, this paper takes Siamese network as the basic structure and proposes a localization method based on similarity learning and scene category filtering, as well as reranking. Firstly, the cross-view images are used for similarity learning, and the reference remote sensing images are sorted according to the similarity values. Then, the categories of ground-view and remote sensing images are obtained by training SVM classifier. Finally, geolocalization is performed using feature similarity and scene category information, as well as reranking. The experimental results show that the proposed method can improve geolocalization performance compared to conventional feature matching-based methods. Moreover, the results also demonstrate the effectiveness of scene category filtering and reranking for geolocalization.

Key words: convolutional neural network, geolocalization, feature similarity, category filtering, reranking