Structure-Aware and Temporally Coherent 3D Human Pose Estimation

Abstract

Deep learning methods for 3D human pose estimation from RGB images require a huge amount of domain-specific labeled data for good in-the-wild performance. However, obtaining annotated 3D pose data requires a complex motion capture setup which is generally limited to controlled settings. We propose a semi-supervised learning method using a structure-aware loss function which is able to utilize abundant 2D data to learn 3D information. Furthermore, we present a simple temporal network which uses additional context present in pose sequences to improve and temporally harmonize the pose estimates. Our complete pipeline improves upon the state-of-the-art by 11.8% and works at 30 FPS on a commodity graphics card.