UAV-based cross-view geo-localization fusion spatial attention mechanism and Netvlad

Abstract

The purpose of cross-view image geo-localization is to retrieve the same geographic target from images with different views acquired from different platforms. Facing the great differences in the appearance of images from different viewpoints, we propose an image retrieval method (spatial-Netvlad-siamese net, SNSnet) that fuses the trainable local aggregation descriptor vector (Netvlad) and the spatial attention mechanism. SNSnet can simultaneously process two different viewpoint images, and achieve image retrieval and matching by increasing the distance between mismatched image pairs and decreasing the distance between matched image pairs. We conducted experiments on mutual retrieval between satellite viewpoint images and UAV viewpoint images and achieved good results.