Human identification from images captured in unconstrained scenarios is still an unsolved problem, which finds applications in several areas, ranging from all the settings typical of video surveillance, to robotics, metadata enrichment of social media content, and mobile applications. The most recent approaches rely on techniques such as sparse coding and low-rank matrix decomposition. Those build a generative representation of the data that on the one hand, attempts capturing all the information descriptive of an identity; on the other hand, training and testing are complex to allow those algorithms to be robust against grossly corrupted data, which are typical of unconstrained scenarios.;This thesis introduces a novel low-rank modeling framework for human identification. The approach is supervised, gives up developing a generative representation, and focuses on learning the subspace of nuisance factors, responsible for data corruption. The goal of the model is to learn how to project data onto the orthogonal complement of the nuisance factor subspace, where data become invariant to nuisance factors, thus enabling the use of simple geometry to cope with unwanted corruptions and efficiently do classification. The proposed approach inherently promotes class separation and is computationally efficient, especially at testing time. It has been evaluated for doing face recognition with grossly corrupted training and testing data, obtaining very promising results. The approach has also been challenged with a person re-identification experiment, showing results comparable with the state-of-the-art.
展开▼
