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; Proceedings of the Winter Conference on Applications of Computer Vision (WACV) Workshops, 2025, pp. 479-488

Static cross-view geo-localization datasets fail to capture the dynamic nature of real-world environments as they do not account for rapid urban development and seasonal changes. As a result models trained on such datasets experience degraded performance when confronted with more recent data as they struggle to adapt to temporal variations such as newly constructed buildings or changing landscapes. To accurately assess the performance gap and evaluate model robustness it is essential to use temporally diverse data that allows us to measure how well models can handle temporal shifts and remain resilient to such changes. To address this need we have enriched the CVUSA dataset with recent satellite and Street View imagery creating the CVTemporal dataset. This enhanced dataset is critical for testing how well geo-localization models can adapt to temporal discrepancies and identify persistent invariant features. In this work we also examine the impact of temporal changes on the performance of selected well-known cross-view geo-localization models.Furthermore we present a re-ranking approach based on existing satellite imagery from both datasets which leads to significant performance improvements. Despite temporal variations in the data we achieve with our models remarkably good results especially on R@1. Additionally we investigate strategies for identifying which temporally changed data should be collected to update pre-trained models minimizing the labor-intensive process of recollecting entire datasets. Experiments with the CVUSA datasets demonstrate that it is possible to improve temporal alignment and model performance with only a small fraction of newly collected data.