Global Cross-Modal Geo-Localization: A Million-Scale Dataset and a Physical Consistency Learning Framework
The CORE dataset provides a million-scale collection of cross-view geo-localized images from 225 regions globally, addressing limitations in existing CMGL research. The PLANET framework is a physical-law-aware network that uses contrastive learning to align text descriptions with intrinsic physical signatures like spectral color and geometric structure in satellite imagery, achieving state-of-the-art global localization accuracy.