SkyPart: Prototype-Based Semantic Part Discovery for Weather-Robust Cross-View Geo-Localization

A new method called SkyPart improves cross-view geo-localization (CVGL) for autonomous drone navigation when GNSS signals are unavailable. CVGL matches oblique drone views to geo-referenced satellite tiles, but existing approaches suffer from compressing patch grids into single vectors without separating layout from texture, retaining altitude-related scale variation, and relying on hand-tuned scalars for multi-objective training. SkyPart is a lightweight swappable head for patch-based vision transformers that institutes explicit part grouping over the patch grid. It features four components: learnable prototypes competing for patch tokens via single-pass cosine assignment, altitude-conditioned linear modulation, and other theory-grounded mechanisms. The method is designed to be robust to weather conditions, addressing a key limitation in prior work. The paper is available on arXiv under identifier 2605.11654.