Wang Tingyu ; Zheng Zhedong ; Yan Chenggang ; Zhang Jiyong ; Sun Yaoqi ; Zheng Bolun ; Yang Yi

Abstract

Cross-view geo-localization is to spot images of the same geographic targetfrom different platforms, e.g., drone-view cameras and satellites. It ischallenging in the large visual appearance changes caused by extreme viewpointvariations. Existing methods usually concentrate on mining the fine-grainedfeature of the geographic target in the image center, but underestimate thecontextual information in neighbor areas. In this work, we argue that neighborareas can be leveraged as auxiliary information, enriching discriminative cluesfor geolocalization. Specifically, we introduce a simple and effective deepneural network, called Local Pattern Network (LPN), to take advantage ofcontextual information in an end-to-end manner. Without using extra partestimators, LPN adopts a square-ring feature partition strategy, which providesthe attention according to the distance to the image center. It eases the partmatching and enables the part-wise representation learning. Owing to thesquare-ring partition design, the proposed LPN has good scalability to rotationvariations and achieves competitive results on three prevailing benchmarks,i.e., University-1652, CVUSA and CVACT. Besides, we also show the proposed LPNcan be easily embedded into other frameworks to further boost performance.