Unified Human-Centric Model, Framework and Benchmark: A Survey

Human-centric Computer Vision Tasks (HCTs) refer to a series of tasks related to the human body, such as Human Pose Estimation, Pedestrian Tracking, Re-Identification (ReID), Human Parsing and Action Recognition, etc. In the past three years, a large number of Human-centric Methods (HCMds) for HCTs have emerged, based on a common assumption that these tasks should share the same underlying semantic structure of the human body, integrating multi-modal information could achieve more powerful functionality or lower computational cost. However, a systematic and comprehensive literature review on this field is still missing, this survey provides a comprehensive review of these works. We first give a clear definition and taxonomy standard, then propose a new taxonomy of HCMds. Next, we discuss key technologies and the path towards Unified Human-centric model (UniHCM). Third, following the new taxonomy, we take a brief review and summary highlights of the representative model, framework, and benchmark for HCTs. Finally, We discuss and analyze the limitations of existing HCMds and suggest possible future research directions.