Cross-modal geo-localization (CMGL) aims to match ground-level text descriptions with geo-tagged aerial imagery, supporting tasks like pedestrian navigation and emergency response. The authors argue that existing datasets are limited by narrow geographic coverage and simple scene diversity, which fails to capture the global variety of architectural styles and topography.

To address this, they introduce CORE, described as the first million-scale dataset for global CMGL. It includes 1,034,786 cross-view images sampled from 225 distinct geographic regions across six continents, with varied environmental conditions and urban layouts. The researchers use large vision-language models (LVLMs) to synthesize high-quality scene descriptions that contain discriminative cues.

For the model, they propose PLANET, a physical-law-aware network. PLANET introduces a contrastive learning paradigm that guides textual representations to capture intrinsic physical signatures of satellite imagery. Experiments across varied geographic regions show that PLANET significantly outperforms state-of-the-art methods, establishing a new benchmark for robust global-scale geo-localization. The dataset and code are to be released.