This task is to estimate 2D or 3D human poses from an image or a video. Human poses are often represented as a set of body joint coordinates. Accurate estimation of human poses can help many applications such as action recognition and augmented reality (AR).
A standard bottom-up model that supports real-time multi-person 2D pose estimation. Compared to top-down approaches which detect humans and perform single-person pose estimation for each detection, this model simultaneously detects and associates human body parts with a set of non-parametric limb representations called Part Affinity Fields (PAFs).