Bridging the Domain Gap in the Transformer Era for Cross View Geo-Localization

Cross-view geo-localization has garnered notable attention in the realm of computer vision, driven by the widespread availability of geotagged datasets and advancements in machine learning techniques. This thesis provides a comprehensive survey of cutting-edge methodologies, techniques, and associated challenges integral to this domain, focusing on feature-based and deep learning strategies. Feature-based methods utilize unique features to establish correspondences across different viewpoints, while deep learning methodologies employ convolutional neural networks to embed view-invariant attributes. Building on this survey, the research selects the TransGeo model for its promising approach to addressing the limitations of CNN-based methods. The experiments incorporate segmentation techniques and Bird’s Eye View (BEV) transformations in TransGeo. The thesis concludes by exploring future research avenues and highlighting the potential applications of cross-view geo-localization in an increasingly interconnected global landscape.