Cross-View Object Geo-Localization via Contrastive Learning for UAV and Satelite images

Le résumé fourni par la source

Cross-view object geo-localization (CVOGL) refers to the precise identification of a ground object within a reference image, provided a query image acquired from a different viewpoint. This application possesses substantial potential across a wide range of domains, including urban planning, environ-mental monitoring, and emergency response. Literature studies on CVOGL employ vanilla CNNs to locate the object with the highest similarity, failing to exploit the potential correlations be-tween different perspectives. To address this limitation, this study introduces contrastive learning to explicitly model the consistency within cross-view images. Furthermore, several novel techniques are introduced to improve localization accuracy, including Spatial refined Contrastive Learning (SR-CL) and Cross-View Attention Module(CVAM). Experimental results demonstrate that the pro-posed method achieves significant advancement over the state-of-the-art, exhibiting a nearly 3% improvement in accuracy as evaluated in the CVOGL benchmark dataset.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.