ARC-Loc: Leveraging Azimuthal Ray Convergence as a Geometric Cue for Direct Cross-View Localization

Cross-view localization (CVL) estimates the pose of a groundimage by matching it to a geo-referenced satellite image. To bridge theextreme viewpoint gap, mainstream pipelines rely on Bird’s-Eye-View(BEV) transformations or 2D-to-3D lifting. However, deriving 3D struc-tures from a single ground image is fundamentally ill-posed, causing thesemethods to endure geometric distortions and computational costs dur-ing 3D lifting or BEV projection. Furthermore, relying on external depthfoundation models to resolve this introduces latency and remains suscep-tible to noisy predictions. In this work, we present a different approachinspired by a human navigation technique called resection, that can per-form direct ground-to-satellite image matching and localization withoutrelying on external depth foundation models. The key insights of ourmethod are that (i) ground keypoints can be translated into azimuthalrays on the satellite map, and (ii) these rays ideally converge at theuser location. Exploiting this geometric constraint through direct line-to-point correspondences, we introduce a minimal Azimuthal Ray Con-vergence (ARC) solver to identify the intersection alongside an ARC lossto optimize the matching network. By eliminating dependencies on com-putationally heavy BEV transformations and external depth foundationmodels, our approach achieves faster, memory-efficient inference, whileits explicit feature matching ensures straightforward compatibility withexisting frameworks. Experiments on VIGOR and KITTI demonstratethat ARC-Loc maintains competitive localization accuracy compared torecent approaches, highlighting its practicality.