Lightweight 2D Human Pose Estimation for Fitness Coaching System

Abstract

In this paper, we describe lightweight 2D human pose estimation for a fitness coaching system. To achieve real-time inference speed on mobile devices, we propose the online pose distillation learning strategy that trains large teacher networks and small student networks simultaneously. The proposed lightweight model requires significantly lower computation as well as shows competitive pose estimation accuracy on the COCO keypoint detection dataset. In addition, the model was fine-tuned on a fitness dataset for the fitness coaching application where various fitness postures were accessed.