Deuser Fabian ; Habel Konrad ; Oswald Norbert

Abstract

Cross-View Geo-Localisation is still a challenging task where additionalmodules, specific pre-processing or zooming strategies are necessary todetermine accurate positions of images. Since different views have differentgeometries, pre-processing like polar transformation helps to merge them.However, this results in distorted images which then have to be rectified.Adding hard negatives to the training batch could improve the overallperformance but with the default loss functions in geo-localisation it isdifficult to include them. In this article, we present a simplified buteffective architecture based on contrastive learning with symmetric InfoNCEloss that outperforms current state-of-the-art results. Our framework consistsof a narrow training pipeline that eliminates the need of using aggregationmodules, avoids further pre-processing steps and even increases thegeneralisation capability of the model to unknown regions. We introduce twotypes of sampling strategies for hard negatives. The first explicitly exploitsgeographically neighboring locations to provide a good starting point. Thesecond leverages the visual similarity between the image embeddings in order tomine hard negative samples. Our work shows excellent performance on commoncross-view datasets like CVUSA, CVACT, University-1652 and VIGOR. A comparisonbetween cross-area and same-area settings demonstrate the good generalisationcapability of our model.