AlphaPose: Whole-Body Regional Multi-Person Pose Estimation and Tracking in Real-Time
Hao-Shu FangJiefeng LiHongyang TangChao XuHaoyi ZhuYuliang XiuYong-Lu LiCewu Lu
Hao-Shu FangJiefeng LiHongyang Tang
...+4
Cewu Lu
Nov 2022
0被引用
62笔记
读论文,拿好礼活动火爆进行中,iPad、蓝牙耳机、拍立得、键盘鼠标套装等你来拿!
摘要原文
Accurate whole-body multi-person pose estimation and tracking is an importantyet challenging topic in computer vision. To capture the subtle actions ofhumans for complex behavior analysis, whole-body pose estimation including theface, body, hand and foot is essential over conventional body-only poseestimation. In this paper, we present AlphaPose, a system that can performaccurate whole-body pose estimation and tracking jointly while running inrealtime. To this end, we propose several new techniques: Symmetric IntegralKeypoint Regression (SIKR) for fast and fine localization, Parametric PoseNon-Maximum-Suppression (P-NMS) for eliminating redundant human detections andPose Aware Identity Embedding for jointly pose estimation and tracking. Duringtraining, we resort to Part-Guided Proposal Generator (PGPG) and multi-domainknowledge distillation to further improve the accuracy. Our method is able tolocalize whole-body keypoints accurately and tracks humans simultaneously giveninaccurate bounding boxes and redundant detections. We show a significantimprovement over current state-of-the-art methods in both speed and accuracy onCOCO-wholebody, COCO, PoseTrack, and our proposed Halpe-FullBody poseestimation dataset. Our model, source codes and dataset are made publiclyavailable at https://github.com/MVIG-SJTU/AlphaPose.