Best Paper Award Winner - GeoCV Workshop at WACV 2025
10 January 2025
Visual geo-localization—the task of determining precise GPS coordinates from images alone—remains a challenging problem in the computer vision community. Traditional models struggle with temporal variations, such as newly constructed buildings, seasonal changes, and evolving landscapes, leading to a decline in accuracy over time. To tackle this issue, we explore methods that enhance a model’s adaptability to these changes.
As part of our collaboration between the University of the Bundeswehr Munich (UniBw M) and the Technical University of Munich (TUM), we present CVTemporal, an extension of the widely used CVUSA dataset. CVTemporal enriches the original data with updated satellite and Street View imagery to evaluate the impact of temporal shifts on geo-localization models. Specifically, we updated the 35,000 satellite and Street View image pairs by collecting new images of the same locations with the latest available timestamps. Further in our study, we evaluate how existing cross-view geo-localization models handle these shifts and propose a re-ranking approach that significantly improves performance despite environmental changes. Additionally, we explore strategies to determine the minimal amount of new data required for effective model updates, reducing the need for full dataset re-creation.
Our findings demonstrate that even with limited data updates, temporal robustness in geo-localization can be substantially improved. We are honored that our work has been accepted for presentation at the GeoCV workshop at WACV 2025 in Tucson, Arizona. This paper received the Best Paper Award at the workshop.
Temporal Resilience in Geo-Localization: Adapting to the Continuous Evolution of Urban and Rural Environments
Fabian Deuser, Wejdene Mansour, Hao Li, Konrad Habel, Martin Werner, Norbert Oswald
