Li Chen ; Lee Gim Hee

Abstract

3D human pose estimation from a monocular image or 2D joints is an ill-posedproblem because of depth ambiguity and occluded joints. We argue that 3D humanpose estimation from a monocular input is an inverse problem where multiplefeasible solutions can exist. In this paper, we propose a novel approach togenerate multiple feasible hypotheses of the 3D pose from 2D joints.In contrastto existing deep learning approaches which minimize a mean square error basedon an unimodal Gaussian distribution, our method is able to generate multiplefeasible hypotheses of 3D pose based on a multimodal mixture density networks.Our experiments show that the 3D poses estimated by our approach from an inputof 2D joints are consistent in 2D reprojections, which supports our argumentthat multiple solutions exist for the 2D-to-3D inverse problem. Furthermore, weshow state-of-the-art performance on the Human3.6M dataset in both besthypothesis and multi-view settings, and we demonstrate the generalizationcapacity of our model by testing on the MPII and MPI-INF-3DHP datasets. Ourcode is available at the project website.