Learning human poses in natural scenes

Abstract

[ACCESS RESTRICTED TO THE UNIVERSITY OF MISSOURI SYSTEM AT AUTHOR'S REQUEST.] The task of human pose estimation in natural scenes is to determine the precise pixel locations of body keypoints. It is very important for many high-level computer vision tasks, including action and activity recognition, human-computer interaction, motion capture, and animation. We cover two different approaches for this task: top-down approach and bottom-up approach. In the top-down approach, we propose a human tracking method called ROLO that localizes each person. We then propose a state-of-the-art single-person human pose estimator that predicts the body keypoints of each individual. In the bottomup approach, we propose an efficient multi-person pose estimator with which we participated in a PoseTrack challenge [11]. On top of these, we propose to employ adversarial training to further boost the performance of single-person human pose estimator while generating synthetic images. We also propose a novel PoSeg network that jointly estimates the multi-person human poses and semantically segment the portraits of these persons at pixel-level. Lastly, we extend some of the proposed methods on human pose estimation and portrait segmentation to the task of human parsing, a more finegrained computer vision perception of humans.