Optimizing Geo-Localization with k-Means Re-Ranking in Challenging Weather Conditions

In this paper, we present our solution to the 2nd Workshop on UAVs in Multimedia, focusing on improving image matching for cross-view geo-localization. Our approach utilizes a Vision Transformer pre-trained using the DINOv2 method, which has been shown to provide robust performance. To address the challenges posed by varying environmental conditions, we integrate synthetic weather augmentations to improve the model's adaptability and reliability under different weather scenarios. In addition, we introduce a simple k-Means re-ranking strategy specifically designed for n:1 matching problems. This method not only improves the accuracy of our localization approach, but also demonstrates versatility by being applicable to various other datasets. Our experimental results on the University-160kWX dataset confirm the effectiveness of our approach. We achieve a Recall@1 of 96.30% and a Recall@5 of 98.84%. These results demonstrate the high accuracy of our system, with R@5 indicating that almost all images are correctly identified, except for cases involving highly ambiguous buildings.