Rethinking on Multi-Stage Networks for Human Pose Estimation
Wenbo Li1,2 Zhicheng Wang1∗ Binyi Yin1 Qixiang Peng1 Yuming Du1,3 Tianzi Xiao1,4 Gang Yu1 Hongtao Lu2 Yichen Wei1 Jian Sun1 1Megvii Inc. (Face++) 2Shanghai Jiao Tong University 3Beihang University 4Beijing University of Posts and Telecommunications 2 The first two authors contribute equally to this work. This work is done when Wenbo Li, Binyi Yin, Qixiang Peng, Yuming Du and Tianzi Xiao are interns at Megvii Research.
Abstract
Existing pose estimation approaches fall into two categories: single-stage and multi-stage methods. While multi-stage methods are seemingly more suited for the task, their performance in current practice is not as good as single-stage methods.
中文速览
多阶段姿态估计网络（MSPN）在设计上天然适合人体姿态估计任务，却一直打不过更简单的单阶段方法，作者深入分析后发现根源在于多阶段方法的诸多设计细节存在缺陷。为此，他们提出了三项改进：用现代ResNet风格的网络替换Hourglass中落后的等宽通道模块、引入跨阶段特征聚合来减少反复上下采样带来的信息损失、以及用"粗到细"的渐进式监督（早期阶段用大高斯核、后期阶段用小高斯核）来逐步精化关节定位精度。在MS COCO和MPII两大标准数据集上，改进后的MSPN均刷新了最优成绩，其中在COCO test-challenge上比2017年冠军高出4.3 AP，同时随阶段数增加性能持续提升、不像单阶段方法那样很快饱和，证明多阶段架构只要设计得当，本质上比单阶段方法更有潜力。
原文 arXiv:1901.00148;中英对照 + 大白话阅读 https://aha.fim.ai/paper/1901.00148v4