DEEP LEARNING FOR HUMAN-CENTRIC IMAGE ANALYSIS

Abstract

Though great progresses have been made in recent years, the performance of human-centric image analysis in real-world scenarios is still far from being satisfactory. In this thesis, we systematically investigate the problem of human-centric analysis based on deep learning, from face recognition for exploring the identity information to human parsing for exploring the fine-grained semantic information. Pose and age variations are the most challenging factors for face recognition in unconstrained scenarios. We start by designing pose-invariant face recognition approaches, which aim to address unconstrained face recognition with large/extreme pose variations, by synthesizing photorealistic faces with different poses and learning pose-invariant facial representations. Thereafter, we propose age-invariant face recognition algorithms, which elegantly learn disentangled facial representations and photorealistic cross-age faces to solve cross-age face recognition. To enable more detailed human-centric analysis, we then introduce a novel strategy for instance-agnostic human parsing, which effectively aggregates multi-scale contextual information and incorporates human pose based constraints to improve the results. Finally, we propose a novel nested adversarial learning strategy to address the challenging instance-level human parsing (multi-human parsing), and construct a new large-scale and fine-grained benchmark dataset to further push the research frontier of human-centric image analysis and crowded scene understanding.