GeoSAM: A Sampling and Attention Mechanism for Cross-View Geo-Localisation

Cross-view geo-localisation aims to match street-view images with satellite images, but this task remains challenging due to significant appearance and viewpoint discrepancies. This paper proposes a novel framework, GeoSAM, which enhances feature alignment and matching accuracy by integrating a dynamic sampling strategy and attention mechanism. GeoSAM employs a dual-branch network with Convolutional Block Attention Modules (CBAM) to strengthen feature extraction for both street-view and satellite-view images. Furthermore, the framework incorporates a dynamic hard negative sampling strategy, combining geographic proximity and visual similarity, to ensure robustness in contrastive learning. Experiments on the CVUSA dataset demonstrate that GeoSAM significantly outperforms existing methods in terms of Top-1 accuracy and recall, offering an efficient and generalisable solution for cross-view geo-localisation tasks.