Enhancing Geo-Localization Feasibility through Diffusion-Generated Satellite Images

In this study, we propose a novel approach to enhance satellite image localization by leveraging OpenStreetMap (OSM) data for semantic guidance. Our method employs the GeoSynth technique to generate realistic satellite images from OSM information, effectively bridging the gap between road backbone maps and actual satellite imagery. We then utilize the GAMa framework for same-view satellite image retrieval, allowing us to match generated images with a comprehensive satellite image database. Experimental results demonstrate that our approach achieves a retrieval $\mathrm{R} \text{@} 1 \%$ of $\mathbf{6 7 \%}$, indicating significant potential for improving localization accuracy based on road backbone maps. This work highlights the effectiveness of combining semantic mapping with advanced image synthesis and retrieval techniques in the domain of geo-localization.