提高人体姿态估计网络效果的弱监督方法  

Weakly supervised method for improving effect of human pose estimation network

作　　者：解至煊 孙晓刚[1] XIE Zhixuan;SUN Xiaogang(Chengdu Institute of Computer Application,Chinese Academy of Sciences,Chengdu Sichuan 610041,China;University of Chinese Academy of Science,Beijing 100049,China)

机构地区：[1]中国科学院成都计算机应用研究所,成都610041 [2]中国科学院大学,北京100049

出　　处：《计算机应用》2020年第S02期48-53,共6页journal of Computer Applications

基　　金：四川省重点研发项目(2018GZ0231)。

摘　　要：监督学习通过从大量具有完整标签的训练示例中学习来构建预测模型,在各领域都取得了巨大的成功,但数据标注过程需要消耗巨大的成本,且标注要求越复杂,标注成本越昂贵。关注人体姿态估计网络的弱监督方法,在骨干网络之后增加高、低分辨率子网络,高分辨率网络通过位置特征的相关性加强多个关节点是一个整体的权重,低分辨率网络使网络在训练过程中可接受图像级粗粒度标签。所提方法仅需增加少量有人或无人两种简单标注的图片,即可让网络在目标场景下快速拟合其分布,达到更好的效果。在室内厂房环境、工地施工现场、驾驶室监控环境等不同场景下测试结果表明,所提方法均能达到不错的效果。Supervised learning builds predictive models by learning from a large number of training examples with complete labels,which has achieved great success in various fields,but the data labeling process requires huge costs,and the more complex the labeling requirements,the more expensive the labeling costs.To resolve this problem in human pose estimation network,the weakly supervised method was used by adding high-and low-resolution sub-networks after the backbone network.The high-resolution network strengthened multiple joint points through the correlation of position feature as an overall weight,the low-resolution networks accepted image-level coarse-grained labels during training.Thus,the proposed method only needed to add a small number of images which just need to simply indicate if anyone is there,so that the network could quickly fit its distribution in the target scene to achieve better results.Test results in different scenarios such as factory environment,construction site and cab environment show that the proposed method can achieve good results.

关 键 词：高分辨率网络 弱监督学习 人体姿态估计 关键点检测 多示例学习 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]