基于支持对挖掘的主动学习行人再识别  

Support pair active learning for person re-identification

作　　者：金大鹏 李旻先[1] Jin Dapeng;Li Minxian(School of Computer Science&Engineering,Nanjing University of Science&Technology,Nanjing 210094,China)

机构地区：[1]南京理工大学计算机科学与工程学院,南京210094

出　　处：《计算机应用研究》2023年第4期1220-1225,1255,共7页Application Research of Computers

基　　金：国家自然科学基金资助项目(62076132);江苏省自然科学基金资助项目(BK20211194)。

摘　　要：基于监督学习的行人再识别方法需要大量人工标注的数据,对于实际应用并不适用。为了降低大规模行人再识别的标注成本,提出了一种基于支持对挖掘主动学习(support pair active learning, SPAL)的行人再识别方法。具体地,建立了一种无监督主动学习框架,在该框架中设计了一种双重不确定性选择策略迭代地挖掘支持样本对并提供给标注者标注;其次引入了一种约束聚类算法,将有标签的支持样本对的关系传播到其他无标签的样本中;最后提出了一种由无监督对比损失和监督支持样本对损失组成的混合学习策略来学习具有判别性的特征表示。在大规模行人再识别数据集MSMT17上,该方法相比于当前最先进的方法,标注成本降低了64.0%,同时mAP和rank1分别提升了11.0%和14.9%。大量实验结果表明,该方法有效地降低了标注成本并且优于目前最先进的无监督主动学习行人再识别方法。Supervised-learning based person re-identification requires a large amount of manual labeled data,which is not applicable in practical deployment.This paper proposed a support pairs active learning(SPAL)re-identification framework to lower the manual labeling cost for large-scale person re-identification.Specifically,this paper built a kind of unsupervised active learning framework,and it designed a dual uncertainty selection strategy to iteratively discover support pairs and required human annotations in this framework.Afterwards,it introduced a constrained clustering algorithm to propagate the relationships of labeled support pairs to other unlabeled samples.Moreover,it proposed a hybrid learning strategy consisting of an unsupervised contrastive loss and a supervised support pairs loss to learn the discriminative feature representation.On large-scale person re-identification dataset MSMT17,compared with the state-of-the-art methods,the labeling cost of the proposed method is reduced by 64%,mAP and rank1 are increased by 11.0%and 14.9%respectively.Extensive experiments demonstrate that it can effectively lower the labeling cost and is superior to state-of-the-art unsupervised active learning person re-identification methods.

关 键 词：行人再识别 无监督主动学习 约束聚类 不确定性选择 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]