Zhou Xingyi Huang Qixing Sun Xiao Xue Xiangyang Wei Yichen

Abstract

In this paper, we study the task of 3D human pose estimation in the wild.This task is challenging due to lack of training data, as existing datasets areeither in the wild images with 2D pose or in the lab images with 3D pose. We propose a weakly-supervised transfer learning method that uses mixed 2Dand 3D labels in a unified deep neutral network that presents two-stagecascaded structure. Our network augments a state-of-the-art 2D pose estimationsub-network with a 3D depth regression sub-network. Unlike previous two stageapproaches that train the two sub-networks sequentially and separately, ourtraining is end-to-end and fully exploits the correlation between the 2D poseand depth estimation sub-tasks. The deep features are better learnt throughshared representations. In doing so, the 3D pose labels in controlled labenvironments are transferred to in the wild images. In addition, we introduce a3D geometric constraint to regularize the 3D pose prediction, which iseffective in the absence of ground truth depth labels. Our method achievescompetitive results on both 2D and 3D benchmarks.