作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2018, Vol. 44 ›› Issue (11): 257-264. doi: 10.19678/j.issn.1000-3428.0050384

• 图形图像处理 • 上一篇    下一篇

面向姿态估计的组件感知自适应算法

陈耀东1a,刘琴2,彭蝶飞1b   

  1. 1.长沙师范学院 a.信息与工程系; b.科研与学科建设处,长沙 410100; 2.湖南省教育科学研究院 职业教育与成人教育研究所,长沙 410005
  • 收稿日期:2018-02-01 出版日期:2018-11-15 发布日期:2018-11-15
  • 作者简介:陈耀东(1978—),男,讲师、博士,主研方向为计算机视觉、机器学习;刘琴,副教授;彭蝶飞,教授、博士。
  • 基金资助:

    教育部人文社科基金青年项目“大数据驱动下学龄前儿童行为问题分析与预测研究”(17YJCZH028);湖南省教育厅科学研究项目优秀青年项目“生态小镇智慧管理平台构建——以浔龙河为个案的研究”(16B025);湖南省社会科学联合会智库项目“基于智慧旅游的景区游客体验模型构建及优化调控研究”(ZK17)。

Component aware Adaptive Algorithm for Pose Estimation

CHEN Yaodong1a,LIU Qin2,PENG Diefei1b   

  1. 1a.Department of Information and Engineering; 1b.Department of Research and Discipline Development, Changsha Normal University,Changsha 410100,China; 2.Department of Vocational and Adult Education Research,Hunan Provincial Research Institute of Education,Changsha 410005,China
  • Received:2018-02-01 Online:2018-11-15 Published:2018-11-15

摘要:

针对姿态估计的结构化输出特点,提出一种领域自适应学习算法。建立一种组件感知的参数学习过程,根据目标的各组件调整自适应参数,提升模型的泛化能力。依据领域自适应算法特点引入基于主动学习样本选取策略,提升模型的学习效率。对特征分布差异的2种训练场景进行模拟实验,结果表明,该算法训练的模型在平均准确率上比传统学习算法提升6%~8%,比已有的自适应算法提升2%~4%,使用样本选取策略后,则进一步提升约2%。

关键词: 领域自适应, 组件感知, 姿态估计, 主动学习, 结构化支持向量机

Abstract: This paper introduces a new domain adaptation algorithm for training pose estimation parameters that are characterized by structured output prediction.The algorithm establishes a component-aware parameter learning process,which can adjust the adaptive parameters according to the components of the target,and is suitable for the scenes with different feature distributions to improve the generalization ability of the pose estimation model.In order to further improve the learning efficiency of the model,an active learning sample selection strategy is introduced according to the characteristics of the domain adaptive algorithm.The experiments simulate two kinds of training scenarios with different feature distributions,i.e.,different pose distribution and different appearance distribution.Experimental results show that the average precision of the proposed algorithm is improved by 6%~8% compared with the traditional learning algorithms and 2%~4%compared with the existing adaptive algorithms,after using the sample selection strategies,it will increase by about 2%.

Key words: domain adaptation, component-aware, pose estimation, active learning, Structural Support Vector Machine(S-SVM)

中图分类号: