Single-Stage UAV Geolocation Enhancement with MaSA and Self-Homologous Loss

Visual Geo-localization plays a pivotal role in the domains of computer vision and image processing. In UAV nav-igation, high-precision localization can be achieved by matching the UAV's current viewpoint image with geo-tagged reference images, especially in scenarios where G PS signals are unavailable. However, conventional two-stage image retrieval methods, which initially employ global features for coarse retrieval followed by local feature-based re-ranking, are computationally intensive and demand substantial storage. To address these limitations, we propose a novel method that integrates explicit spatial priors from the Manhattan Self-Attention (MaSA) mechanism with the robust global modeling capabilities of Transformers, thereby streamlining the retrieval process into a single-stage. Furthermore, leveraging the unique pose characteristics of UAV-captured images, we refine the loss function and introduce a UAV-specific self-homogenous triplet loss to enhance the robustness of global feature extraction. Additionally, we have developed a new visual localization dataset comprising UAV and satellite images and validated our approach on this dataset. Experimental results demonstrate that our method outperforms state-of-the-art techniques, achieving improvements of 3.92% and 1.81% in top-1 and top-S retrieval accuracy, respectively.