基于分布约束的非对称度量学习无监督行人重识别
DOI:
作者:
作者单位:

山东理工大学 电气与电子工程学院

作者简介:

通讯作者:

中图分类号:

TP391

基金项目:

国家自然科学基金(61801272);淄博市张店区校城融合项目(118228);山东省自然科学基金项目(ZR2015FL029, ZR2016FL14)


Asymmetric metric learning approachbased on distribution constraints for unsupervised person re-identification
Author:
Affiliation:

School of Electrical and Electronic Engineering,Shandong University of Technology

Fund Project:

The National Natural Science Foundation of China(No.61801272);The Integration Funds of Shandong University of Technology and Zhangdian District (No.118228);The Natural Science Foundation of Shandong Province of China (No. ZR2015FL029, ZR2016FL14)

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
  • |
  • 文章评论
    摘要:

    在无监督行人重识别中,针对传统非对称度量学习方法无法克服不同视角的数据分布差异问题,提出一种基于分布约束的非对称度量学习无监督行人重识别方法。首先,采用JSTL技术对特征提取网络预训练,得到具有较强鲁棒性的特征表示;然后,提出基于分布约束的非对称度量学习算法,通过在传统非对称度量学习目标函数中引入分布约束,实现不同摄像视角下行人图像非对称特征变换的同时,有效克服了行人数据分布差异导致的识别精确度低的问题;最后,采用梯度下降法优化目标函数,并通过广义特征值问题求解获得最优度量矩阵。基于Market和Duke两个公共数据集的实验表明,该算法的rank1值分别达到57.01%和32.32%,map值分别达到27.91%和16.00%,与传统非对称度量学习算法相比识别性能有明显提升。

    Abstract:

    In the field of unsupervised person re-identification, aiming at the problem that the traditional asymmetric metric learning method cannot overcome the difference of data distribution from different views, an asymmetric metric learning approach based on distribution constraints for unsupervised person re-identification is proposed in this paper. First, JSTL technology is used to pre-train the feature extraction network to obtain a robust feature representation. Then, the asymmetric metric learning method based on distribution constraints is proposed. By introducing distribution constraints into the traditional asymmetric metric learning objective function, this method not only realizes the asymmetric feature transformation of person images from different camera views, but also effectively overcomes the problem of low recognition accuracy caused by the difference of person data distribution. Finally, the objective function is optimized by using the gradient descent method, and the optimal measure matrix is obtained by solving the generalized eigenvalue problem. Experiments are implemented on Market and Duke datasets, and the results show that the rank1 value of the algorithm is 57.01% and 32.32%, the map value is 27.91% and 16.00%, respectively, and the recognition performance of the algorithm is significantly improved compared with the traditional asymmetric metric learning algorithm.

    参考文献
    相似文献
    引证文献
引用本文
分享
文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:2021-09-13
  • 最后修改日期:2022-02-27
  • 录用日期:2022-03-15
  • 在线发布日期:
  • 出版日期: