OriLoc: Unlimited-FoV and Orientation-Free Cross-View Geolocalization

Cross-view image-based geolocalization enables accurate, drift-free navigation without external positioning signals, crucial for UAV delivery and disaster relief. However, existing research primarily focuses on ground panoramic images with known orientations, while real-world scenarios involve unknown orientations and limited field of view (FoV), creating a research-application gap. We introduce OriLoc, an innovative cross-view geolocalization method integrating sophisticated orientation estimation for limited FoV and arbitrary orientation scenarios. Our approach employs a dual-weighted soft-margin triplet loss with hard sample mining to extract discriminative features. Additionally, we develop an orientation estimation module using convolution-based sliding windows to assess similarity between satellite-view and query embeddings. The method demonstrates superior performance on three challenging datasets spanning commercial, residential, urban, and suburban areas across two continents. Results show that hard sample mining combined with appropriate learning objectives significantly enhances geolocalization for limited FoV and orientation-free images. Our orientation estimation module achieves remarkable accuracy when integrated with attention embeddings prior to polar transformation. Code and trained models are available at https://github.com/bonihu/OriLoc. Fig. 1.

Cross-view geolocalization performance comparison in different FoV and orientation-free scenarios. (a) Recall@1, (b) Recall@5, (c) Recall@10, and (d) Recall@1% accuracy on the CVUSA benchmark. The angles $70^\circ$, $90^\circ$, $180^\circ$, and $360^\circ$ represent various FoV of unknown direction ground-view images, and “known” indicates that the orientation of the ground-view images is known. Our OriLoc demonstrates superior overall performance.

Furthermore, our orientation estimation module, which incorporates classical convolutions, exhibits remarkable accuracy in estimating orientations when integrated with attention embeddings prior to polar transformation. Code and trained models are publicly available on https://github.com/boni-hu/OriLoc.