Cross-view Geo-localization Based on Cross-domain Matching

As a recently emerging problem, cross-view geo-localization aims at finding image pairs captured from different views (e.g., drone and satellite views) or domains yet same location, which can be widely employed in various applications. However, unlike traditional scene classification problem, it faces several challenges, including large intra-class distance and small inter-class distance caused by domain gap, as well as redundant contextual information and visual distractors across views. To address the concerns, we propose a novel cross-domain matching framework to handle this task, which measures the similarity for query and candidate images from two different domains. Comparing to prior classification based framework, our matching based framework is better suited for the task by forcing the model to learn discriminative features for scenes. Moreover, to aid cross-domain matching, we propose a matching-oriented feature modulation scheme, in which we not only apply a large-view attention module to enhance spatial features but also employ channel shuffling to loose the correlation of key feature semantics and distractors in the respective domains. Last, we conduct experiments to show that our model achieves the state-of-the-art performance and surpasses the competing method by a large margin on the public benchmarks.