Self-supervised learning for fully unsupervised re-identification in real-world applications

Re-Identification (ReID) is vital for real-world applications such as AI-powered security, event understanding, and smart city development. It aims to retrieve all instances of a given person or object across a network of non-overlapping cameras, based solely on visual appearance. It is challenging due to occlusions, viewpoint changes, and background similarities. Supervised methods perform well but rely on costly, biased annotations, limiting scalability. To address this, we propose self-supervised learning algorithms for Unsupervised ReID, extendable to other modalities, like Text Authorship Verification, marked by high intra-class variation and low inter-class distinction. Our work introduces three fully unsupervised ReID methods: one using camera labels, one without side information, and one scalable to larger datasets. We also present a fourth hybrid method for long-range recognition under distortions. These solutions enhance generalization and enable real-world applications in forensics and biometrics. We have released open-source code and demonstrated practical impact, including a consultancy project for the Public Prosecutor’s Office of the State of São Paulo (MPSP), and House of Representatives (Brazilian Federal Chamber and Senate). This research was recognized by the Brazilian Computing Society (SBC) as the best Ph.D. thesis defended in Brazil in 2024.

Conference paper