计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (18): 195-204.DOI: 10.3778/j.issn.1002-8331.2102-0013

• 模式识别与人工智能 • 上一篇    下一篇

融合多聚类信息的无监督行人重识别算法

苏荻翔,王帮海,叶子成   

  1. 广东工业大学 计算机学院,广州 510006
  • 出版日期:2022-09-15 发布日期:2022-09-15

Unsupervised Person Re-Identification Algorithm by Fusing Multi-Cluster Information

SU Dixiang, WANG Banghai, YE Zicheng   

  1. School of Computer, Guangdong University of Technology, Guangzhou 510006, China
  • Online:2022-09-15 Published:2022-09-15

摘要: 在无监督的行人重识别领域中,始终很难对数据集中的难样本对进行很好的挖掘。针对这个问题,提出了融合多种聚类信息生成软多重标签并进行难样本对挖掘的方法。该方法基于不同聚类方法使用的聚类机制不同这一原理,发掘类内样本的共通性与类间样本的差异性,进而使模型能够学习到更有区分性的特征。在Market-1501数据集与DukeMTMC-reID数据集上进行的对比实验结果表明,提出的方法在原来初步学习的网络的基础上,mAP分别提高了14.4%与8.9%,精度均提高显著。

关键词: 行人重识别, 神经网络, 聚类, 难样本挖掘

Abstract: It is never easy to conduct satisfactory data mining in the feild of unsupervised person re-identification. Aiming at this problem, a method that integrates multiple types of clustering information is proposed to generate soft multi-labels and mine hard sample pairs. Based on the principle that different clustering methods use different clustering mechanisms, this method explores the commonality of samples within a class and the differences of samples between classes, thus the model can learn more distinguishing features. The results of comparative experiments carried out on the Market-1501 and the DukeMTMC-reID show that, compared with the original network, the mAP of the proposed method is increased by 14.4% and 8.9% respectively, indicating significant accuracy enhancement.

Key words: person re-identification, neural network, cluster, hard samples mining