LocaMage: Cross-Modal Geolocalization Using Satellite and Ground Imagery

Image geolocalization — the task of determining where a photo was taken based solely on its visual content — remains a significant challenge in computer vision. We introduce LocaMage, a novel approach that leverages both ground-view and satellite imagery processed using the CLIP (Contrastive Language-Image Pre-training) model to improve geolocalization accuracy. Our method addresses key limitations of existing approaches by creating composite inputs of vertically stacked ground-view and satellite imagery. We then encode these inputs using CLIP’s image encoder and correlate the resulting embeddings with geographical locations through a classification and refinement strategy. We implemented image augmentation techniques to improve model generalization. Experiments on the OpenStreetView-5M dataset demonstrate that LocaMage achieves state-of-the-art performance with a mean error of 246 km (compared to 318 km for state-of-the-art methods), a median error of 51 km, and 72% accuracy within 100 km. Additionally, LocaMage attains an average GeoGuessr score of 4502, surpassing both state-of-the-art methods including PIGEON and professional human players. These results highlight the effectiveness of integrating satellite imagery with ground-view images processed through a robust foundation model like CLIP for accurate, planet-scale geolocalization.