Aprendizado auto-supervisionado para re-identificação totalmente não-anotada em aplicações no mundo real

One of the most complex problems in Machine Learning is dealing with unlabeled data. Most top-ranking models rely on massive labeled data to achieve state-of-art results. However, data labeling is not easy nor reliable to obtain due to the highly time-consuming, costly, and error-prone task of annotation. Moreover, bias in the labeled data might be propagated to the model, hindering its performance and generalization. It is paramount to develop methods that can mine patterns in a fully-unsupervised scenario allowing a fast and bias-alleviated deployment. These models could be used in a range of applications, such as forensic investigations, biometrics, and event understanding. This research proposes self-supervised learning algorithms to deal with unlabeled data for deployment in challenging label-absent scenarios. A challenging setup might contain high intra-class disparity (features from the same class are far away from each other in the feature space) and high inter-class similarity (samples from different classes might be closer to each other). To instantiate this complex requirement with applications that capture the mentioned challenges, our exploration focuses on two applications: Unsupervised Re-Identification (ReID) of People and Objects, due to their applicability to event understanding, and on the Text Authorship Verification task. Considering these applications, in this thesis, we propose four methods that deal with varied levels of complexity in unsupervised scenarios. Our first three solutions target the Unsupervised Person ReID task where we assume we do not have identity labeling, i.e., we do not know "who" is detected in the image. The first solution considers meta-information, such as camera labels, to effectively address the task. As there are scenarios where it is not applicable, our second solution is fully unsupervised, i.e., it does not require any side information. Because of this, it can be applied to further tasks than Person ReID in different modalities, such as Text Authorship Attribution in social media posts. The third method also deals with fully unsupervised re-identification scenarios but in large-scale datasets. We also show that this solution can be applied to object re-identification, specifically vehicles. The fourth solution changes the setup by considering supervised training, however targeting long-range recognition. It learns from images mainly distorted by atmospheric turbulence and achieves state-of-the-art results in both Person ReID and Face Recognition tasks. The proposed solutions can be implemented as part of forensic and biometrics pipelines. For instance, they can be employed for event understanding where authorities aim to find possible suspects and investigate people's behavior as well as their possible relationships with objects in a scene. They can be used to get an understanding of what happened and possible investigation insights. The solutions can be also employed in AI-powered biometrics for security-sensitive protection in places such as government facilities, border security, critical infrastructure, and counterterrorism (AU)