Occlusion-robust Cross-view Geo-image Retrieval via Contrastive Learning and Random Masking

Cross-view geo-image retrieval (CVIR) aims to retrieve geographically corresponding satellite images given ground-view images, playing a vital role in various applications such as autonomous driving and unmanned aerial vehicles (UAVs). Although prior studies have mainly focused on mitigating the viewpoint gap through geometric transformation techniques, they have largely overlooked real-world challenges such as occlusions caused by dynamic objects and infrastructure. To address this issue, we propose a contrastive learning framework that introduces random masking to effectively simulate diverse occlusion scenarios during training and enhance the model’s robustness. By applying random masking to the input images during training, the proposed method reduces the model’s reliance on local visual features and enables it to learn more robust features. Our approach requires no transformation or alignment processes, providing a straightforward and effective solution for the occlusion gap. Extensive experiments on standard benchmarks demonstrate that the proposed framework achieves state-of-the-art performance while maintaining robustness under various occlusion conditions.