TOWARDS REALISTIC HUMAN ANALYTICS

Abstract

Human analytics is one of the fundamental research directions in the realm of computer vision. It includes various human-oriented vision tasks, such as the analysis, recognition and synthesis of face and human body. Unlike most previous works under the constrained environment, applications in the realistic environment suffer from scale issues, complex environment and aging of the human. In spite of the enormous efforts within the field, these problems have not been well addressed. There is still a considerable gap between current academic progress and practical applications in the realistic uncontrolled environments.

In this thesis, we focus on addressing some of the key challenges towards realistic human analytics, including datasets in the unconstrained environment, low quality surveillance, scale issue, temporal change caused aging and artificial makeover. To be more specific, several ``in the wild" datasets are constructed, which are carefully designed in consideration of these issues. These datasets are considered as benchmarks and made publicly available to boost the research in these directions. Several algorithms are also proposed and evaluated on these benchmarks which provide considerate improvements over the state-of-the-art approaches.