Beyond Static: Lifelong Cross-View Geo-Localization in Dynamic Environments
Abstract
Cross-View Geo-Localization (CVGL) supports a wide range of geography applications by establishing precise correspondences between low-altitude platforms and satellite imagery. Despite recent performance gains with advances in visual representation learning, current research remains confined to static settings, neglecting real-world demands for streaming data and dynamic environmental evolution (e.g., urban renewal). To address this, we introduce a pioneering yet practical task termed Lifelong Cross-View Geo-Localization (L-CVGL), which requires continuous adaptation to new domains without forgetting prior knowledge. Specifically, compared to standard lifelong learning, L-CVGL faces two unique challenges in balancing stability and plasticity. First, maintaining stability is severely hindered by inherent visual ambiguity, which creates a densely populated feature space where naive adaptation inevitably induces feature aliasing and destroys historical integrity. Second, promoting plasticity is inherently constrained by the underlying retrieval mechanism, as the strict necessity to preserve relational consistency renders conventional rigid anti-forgetting paradigms ineffective for modeling new domains. To address aforementioned challenges, we propose an innovative Omni-view Subspace-Topological Knowledge Transfer (OST-KT) framework tailored to L-CVGL. Specifically, OST-KT integrates Omni-View Subspace Stability Transfer to preserve historical integrity via imposing constraints in the principal subspace, and On-the-Fly Topological Plasticity Transfer to facilitate flexible adaptation through relational topology alignment. To harmonize these modules, we introduce Dynamic Reliability Calibration, which employs a dual-gating strategy to mitigate negative transfer and ensure reliable supervision. Extensive experiments demonstrate that OST-KT significantly mitigates catastrophic forgetting while effectively adapting to new domains, consistently outperforming state-of-the-art methods.