GeoMetric: Injecting Metric Geographic Structure into Worldwide Image Geo-Localization
GeoMetric is a retrieval-based framework for worldwide image geo-localization that encodes GPS coordinates relationally by injecting distance structure into learning and inference. It uses a Transformer encoder with distance-aware attention to ensure spatially proximate samples contribute more strongly, leading to up to 1.5% improvement in street-level accuracy by explicitly modeling spatial proximity.