Deep learning for Sequence-based Visual Geo-localization

Visual geo-localization is the task of recognizing the geographical location where an image was taken by comparing it to a database of geo-tagged images of previously visited places. This is an open topic of research in computer vision, with numerous recent studies that investigate solutions to make these methods more accurate, robust to appearance changes, generalizable and scalable. This thesis proposes to address two emerging problems in visual geo-localization: - the first aim is to perform an extensive survey of the existing methods in the literature for such task, in order to 1) establish a fair and clear evaluation protocol of these methods, which is missing for visual geo-localization, and 2) provide guidelines on the most suitable approach for practical application. - the second contribution, more research oriented, is to assess the extension of the state-of-the-art methods for visual geo-localization from a single image to the case where the input is a sequence of frames. This is a setting that is particularly meaningful for autonomous vehicles or for augmented reality applications. This second part of the thesis explores the possibility to tackle the sequence-based problem using Transformers-based methods, a solution that has not yet been studied in literature.