Object-Level Explanations for Image Geolocation Models: a GeoGuessr use-case

Researchers developed an object-centric attribution pipeline that segments Grad-CAM heatmaps into discrete visual entities and validates their predictive relevance through deletion/insertion tests, revealing whether geolocation models rely on human-interpretable cues like road markings and architecture rather than diffuse spatial patterns.