Image retrieval for Visual Localization and Geo-Localization beyond standard domains: dealing with domain shift in large-scale datasets and challenging indoor environments.

Domain Adaptation (DA) is the task of making a model function effectively on domains that differ from the one(s) it was trained on. Visual (Geo-)Localization, on the other hand, is the task of estimating the position or the pose in which a query image has been taken by comparing it to a large labeled database. The objective of this thesis is to investigate two scenarios where this paradigm is deployed in challenging domains which, without proper countermeasures, would cause severe degradation in performances. In particular, the first one regards the deployment of a large scale model in a different city from the one that it has been trained on. This is done through a carefully picked domain adaptation strategy that has then been accurately tuned to best fit the problem. The second scenario questions the performances of state of the art methods in complex indoor environments and presents a simple yet effective fine-tuning procedure that significantly boosts their performances of all models across all the relevant benchmark.