| Recent years have witnessed an increasing trend of research on monocular 3D human pose estimation because of its wide applications in augmented reality,human-computer interaction,and video analysis.This paper aims to address the problem of estimating absolute 3D poses of multiple people simultaneously from a single RGB image by deep learning,which is different from most of the existing researches that only focus on the estimation of the relative 3D pose of single person.The absolute multi-person 3D pose in this paper includes not only the relative 3D pose of human body,but also the absolute 3D position of human body in the camera coordinate system.That is to say,the absolute 3D coordinates of the keypoints of human bodies in the camera coordinate system.The main achievements of this paper are:1.This paper proposes a novel bottom-up approach to multi-person 3D pose estimation,which predicts 2D poses,absolute 3D positions,and poses of multiple people in a single forward pass.A multi-task fully convolutional network is designed to simultaneously regress the root depth map,2D keypoint heatmaps,part affinity fields(PAFs),and part relative-depth maps.Then,the detected 2D keypoints are grouped into individuals based on PAFs using a part association algorithm and lifted to 3D based on the root depth map and part relative-depth maps.2.This paper proposes a micro network to complement and optimize the 3D human pose estimation results,and a novel depth-aware part association algorithm to solve the ambiguity problem of 2D part association.The effectiveness of these methods is proved by ablation experiments.3.This paper also shows that predicting depths of human bodies is beneficial for 2D and 3D pose estimation,and realizes the whole pipeline of depth estimation task from collecting data to training network with a new random vector loss.Experiments show that the proposed approach achieves state-of-the-art performance on public benchmarks and is generalizable to in-the-wild images. |
PDF Full Text Request