A coarse-to-fine framework of cross-view geo-localization for restricted field of view ground imagery using orientation prediction and variable pixel-position embedding

Cross-view geo-localization aims to localize ground images by matching them with Remote Sensing (RS) images in large geo-annotated databases. It faces significant challenges when ground images have a restricted Field of View (FoV) and unknown orientation, a common scenario in real-world engineering applications such as autonomous driving and drone navigation in GNSS-denied environments. Existing methods lack explicit orientation modeling for restricted FoV images. To address this issue, we propose GeoMatch4RFoV, a novel coarse-to-fine framework of cross-view geo-localization for restricted field of view ground imagery using orientation prediction and variable pixel-position embedding. To ensure computational efficiency suitable for real-world applications, we adopt VMamba as a lightweight backbone for feature extraction. The core innovation of GeoMatch4RFoV lies in handling the restricted FoV challenge: First, an Orientation Prediction Module (OPM) is proposed to align the orientation of the ground image within the RS image, to explicitly capture the geometric misalignment, providing robust orientation cues even under severely restricted FoV settings. Second, a Pixel-Position Embedding-based Feature Aggregation Module (PE-FAM) is designed to generate orientation-adaptive positional embeddings. Unlike static encodings, PE-FAM dynamically recalibrates the spatial context of RS imagery based on the predicted heading, guiding the model to attend to the geometrically corresponding regions for precise feature aggregation. Third, a coarse-to-fine handling strategy is designed to mitigate the significant computational burden imposed by orientation prediction and feature alignment processes. Extensive experiments on two open benchmarks demonstrate that, the proposed model significantly outperforms several state-of-the-art methods.