Self-supervised learning for biometrics and beyond

Abstract

A fundamental problem in machine learning is dealing with unlabeled data. Most models rely on massive labeled data to achieve state-of-art results. However, data labeling is a time-consuming, costly, biased- and error-prone task. It is paramount to develop models that can mine patterns in a fully-unsupervised scenario allowing a fast and bias-free deployment. This project aims at devising self-supervised learning algorithms to deal with unlabeled data for generic label-absent scenarios with challenging setups. A challenging setup might contain high intra-class disparity (features from the same class are far away from each other) and high inter-class similarity (features from different classes might be closer). To instantiate this complex requirement, we explore the Long-range Semi-supervised and Unsupervised Re-Identification of people problem, where images of people are captured from different distances and moments in time, resulting in changes in resolution and appearance. The solution can be extended to objects and places, and used for a range of social impact applications, such as investigations and event understanding. (AU)