Residual Pose: A Decoupled Approach for Depth-based 3D Human Pose Estimation

We propose to leverage recent advances in reliable 2D pose estimation with\nConvolutional Neural Networks (CNN) to estimate the 3D pose of people from\ndepth images in multi-person Human-Robot Interaction (HRI) scenarios. Our\nmethod is based on the observation that using the depth information to obtain\n3D lifted points from 2D body landmark detections provides a rough estimate of\nthe true 3D human pose, thus requiring only a refinement step. In that line our\ncontributions are threefold. (i) we propose to perform 3D pose estimation from\ndepth images by decoupling 2D pose estimation and 3D pose refinement; (ii) we\npropose a deep-learning approach that regresses the residual pose between the\nlifted 3D pose and the true 3D pose; (iii) we show that despite its simplicity,\nour approach achieves very competitive results both in accuracy and speed on\ntwo public datasets and is therefore appealing for multi-person HRI compared to\nrecent state-of-the-art methods.\n