Computer Engineering and Applications ›› 2024, Vol. 60 ›› Issue (1): 227-235.DOI: 10.3778/j.issn.1002-8331.2207-0469

• Graphics and Image Processing • Previous Articles     Next Articles

Unsupervised Person Re-Identification Based on Quadratic Clustering

XIONG Mingfu, XIAO Yingxiong, CHEN Jia, HU Xinrong, PENG Tao   

  1. 1.School of Computer Science and Artificial Intelligence, Wuhan Textile University, Wuhan 430200, China
    2.School of Cyber Science and Engineering, Wuhan University, Wuhan 430072, China
  • Online:2024-01-01 Published:2024-01-01

二次聚类的无监督行人重识别方法

熊明福,肖应雄,陈佳,胡新荣,彭涛   

  1. 1.武汉纺织大学 计算机与人工智能学院,武汉 430200
    2.武汉大学 国家网络安全学院,武汉 430072

Abstract: In view of the influence of objective factors such as hardware differences and illumination changes, the current unsupervised person re-identification method leads to a large contrast in the image of the same person, which is easy to cause the problem of wrong pseudo-labels generation of samples, which makes the existing unsupervised person re-identification method. There is still room for further improvement in the identification method. To solve this problem, this paper proposes an unsupervised person re-identification based on quadratic clustering method. This method mainly includes global quadratic clustering module and supervised learning module based on quadratic clustering results. Specifically, the former performs unsupervised analysis of camera ID and pedestrian ID based on global quadratic clustering, which solves the problem of unified imaging style of the same pedestrian under different camera perspectives; the latter uses supervised learning to improve memory. The initialization and update method of the dictionary solves the problem of model offset during training. Through the co-training of this dual module, it can jointly suppress the problem of false labels generated by images collected across cameras. The algorithm proposed in this paper is tested on Market-1501, DukeMTMC-ReID, MSMT17, Person and VeRi-776 datasets, respectively, and achieves mAP=81.2% and rank-1=91.2%, mAP=68.4% and rank-1=78.7%, mAP=31.1% and rank-1=60.4%, mAP=88.3% and rank-1=93.6%, compared with the current state-of-the-art methods, they have improved by 2.4, 1.8, 6.0, 2.5 and 4.3 percentage points rank-1 accuracy.

Key words: person re-identification, unsupervised learning, quadratic clustering, collaborative training

摘要: 针对当前无监督行人重识别方法因受到硬件差异、光照变化等客观因素的影响,导致同一行人图像出现较大反差,随之易带来样本错误伪标签生成的问题,使得现有无监督行人重识别方法还有待进一步提升的空间。为了解决此问题,提出了一种基于二次重聚类的无监督行人重识别(unsupervised person re-identification based on quadratic clustering)方法。该方法主要包括全局二次聚类的无监督学习模块和基于聚类结果的有监督学习模块。具体来说,前者基于全局二次聚类分别对相机ID和行人身份ID进行无监督分析,解决了同一行人在不同摄像机视角下的统一成像风格问题;后者则采用有监督学习方式改进了内存字典的初始化与更新方式,解决了模型在训练中偏移的问题。通过此双模块的协同训练以共同抑制跨摄像头间采集的图像所产生错误伪标签的问题。所提出的算法分别在Market-1501、DukeMTMC-ReID、MSMT17、Person和VeRi-776数据集上进行实验,取得了mAP=81.2%和rank-1=91.2%、mAP=68.4%和rank-1=78.7%、mAP=31.1%和rank-1=60.4%、mAP=88.3%和rank-1=93.6%的性能,对比当前最先进的方法,分别提高了2.4、1.8、6.0、2.5和4.3个百分点的rank-1准确率。

关键词: 行人重识别, 无监督学习, 二次聚类, 协同训练