HyperPose

Estimating human pose is a core task in many multimedia applications. To fully achieve its promise, users often need to customise pose estimation systems for best possible accuracy, and optimise the systems so that they can achieve real-time processing of high-resolution video streams. To meet these needs, we introduce HyperPose, a library for building pose estimation systems. HyperPose provides a large collection of high-level APIs to help users develop pose estimation algorithms that can achieve high accuracy in the wild. HyperPose further provides a high-performance algorithm execution engine. This engine has a high-performance dataflow for executing pose estimation algorithms. It dynamically dispatches dataflow operators onto CPUs/GPUs, which maximises hardware efficiency, thus achieving real-time processing. Evaluation result show that HyperPose allows users to declare many useful pose estimation algorithms. It also out-performs the performance of state-of-the-art pose estimation systems by up to 3.1x.