基于特征知识蒸馏的人体姿态估计  

Human Pose Estimation Based on Feature Knowledge Distillation

作　　者：袁泽昊 赵嘉莹 YUAN Zehao;ZHAO Jiaying(College of Computer Science and Engineering,Shandong University of Science and Technology,Qingdao Shandong 266000)

机构地区：[1]山东科技大学计算机科学与工程学院,山东青岛266000

出　　处：《软件》2020年第12期198-207,共10页Software

基　　金：国家重点研发计划资助项目(No.2017YFC0804406);山东省重点研发计划资助项目(No.2016ZDJS02A05)。

摘　　要：现有的姿态估计只考虑如何提高模型的识别准确率,却忽略模型结构复杂和计算参数较多而无法在资源受限设备上高效运行的问题。针对这一问题,提出一种基于特征知识蒸馏的人体姿态估计模型。首先构建一种轻量级的高分辨率网络用于学习姿态信息。其次,提出一种新的知识蒸馏方法——特征蒸馏,与姿态蒸馏相结合,分别提取特征图中的高级语义信息和教师网络输出中的“软”知识,从不同角度指导学生网络的训练。实验结果表明,利用该方法得到的学生网络在MPII和COCO数据集上相比知识蒸馏所得到的学生网络识别准确率分别提升了0.6%和1.0%,计算成本却没有明显的增加。表明该方法能够有效的提高识别准确率,具有较高的成本效益。Existing pose estimation only considers how to improve the recognition accuracy of the model,but ignores the problem that the model structure is complex and the calculation parameters are too large to run efficiently on resource-constrained equipment.To solve this problem,a human pose estimation model based on feature knowledge distillation is proposed.First build a lightweight high-resolution network for learning pose information.Secondly,a new method of knowledge distillation,feature distillation,is combined with gesture distillation to extract high-level semantic information in feature maps and"soft"knowledge in teacher network output to guide students'network training from different angles.The experimental results show that the student network obtained by this method has improved the recognition accuracy of the student network on the MPII and COCO datasets by 0.6% and 1.0%,respectively,but the computational cost has not increased significantly.It shows that this method can effectively improve the recognition accuracy and has a high cost-effectiveness.

关 键 词：高分辨率网络 知识蒸馏 人体姿态估计 教师网络 学生网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]