CAFR: Context-Aware Feature Refinement for Narrow Field-of-View Cross-View Geo-Localization

Abstract

While existing work in cross-view geo-localization (CVGL) predominantly localizes 360° panoramic images using satellite imagery, real-world query images often have a narrow field-of-view (NFoV). This narrower field-of-view (FoV) results in a significantly smaller overlapping region between the query and the satellite image, posing a distinct challenge. In narrow CVGL (NCVGL), existing methods lack effective supervision mechanisms to guide models toward overlapping regions, rendering them vulnerable to nonoverlapping content. To address this, we propose context-aware feature refinement (CAFR). Its core innovation lies in a novel feature learning approach: it converts geometric priors within overlapping regions into differentiable supervision signals that directly constrain the cross-attention matrices of features, thereby steering representation learning. CAFR incorporates two key modules: 1) salient region separation (SRS), enforcing spatial consistency between the attention matrices and overlapping regions through a novel saliency separation loss and 2) coordinate-weighted regression (CWR), aligning attention peaks with region centers via a regression loss to enhance feature discriminability. To evaluate model performance under more realistic conditions, we collected user-uploaded real-world images to construct the cross-view narrow FoV (CVNAF) dataset. It is the first NCVGL benchmark that contains perspective images and establishes projection relationships between ground and satellite imagery, offering a realistic evaluation setting for follow-up research. Extensive experiments show that CAFR achieves state-of-the-art (SOTA) performance, improving R@1 by over 5% on multiple NCVGL benchmarks. The code and dataset are released at https://github.com/Hobe-Frank/CAFR