Data-efficient representation learning for visual place recognition

This thesis investigates the problem of visual place recognition, which is a fundamental part of many visual-based localization systems, and therefore of high value to the computer vision community. In particular, we address two problems in the field: the first half of this dissertation is devoted to the presentation and evaluation of a novel dataset in a very under-explored type of environment, namely garden environments. The second half of this thesis addresses the algorithmic part of visual place recognition and proposes a shift of paradigm to learn visual descriptors that encode stronger, reliable, and quantifiable representations of image similarity.