Clustering and Dynamic Sampling Based Unsupervised Domain Adaptation  for Person Re-Identification

CORC > 自动化研究所 > 中国科学院自动化研究所 > 模式识别国家重点实验室 > 生物识别与安全技术研究中心

	Clustering and Dynamic Sampling Based Unsupervised Domain Adaptation for Person Re-Identification
	Wu JL(吴锦林)3,4; Liao SC(廖胜才)2; Wang XB(王晓波)1; Yang Y(杨阳)3,4; Lei Z(雷震)3,4; Li ZQ(李子青)3,4
	2019
会议日期	2019-10
会议地点	中国-上海
英文摘要	Person Re-Identifification (Re-ID) has witnessed great im provements due to the advances of the deep convolutional neural networks (CNN). Despite this, existing methods mainly suffer from the poor generalization ability to unseen persons because of the different characteristics between dif ferent domains. To address this issue, a Clustering and Dy namic Sampling (CDS) method is proposed in this paper, which tries to transfer the useful knowledge of existing la beled source domain to the unlabeled target one. Specififically, to improve the discriminability of CNN model on source do main, we use the commonly shared pedestrian attributes (e.g., gender, hat and clothing color etc.) to enrich the informa tion and resort to the margin-based softmax (e.g., A-Softmax) loss to train the model. For the unlabeled target domain, we iteratively cluster the samples into several centers and dy namically select informative ones from each center to fifine tune the source-domain model. Extensive experiments on the DukeMTMC-reID-reID and Market-1501 datasets show that the proposed method largely improves the state of the arts in unsupervised domain adaptation.
会议录	2019 IEEE International Conference on Multimedia and Expo (ICME)
内容类型	会议论文
源URL	[http://ir.ia.ac.cn/handle/173211/48971]
专题	自动化研究所_模式识别国家重点实验室_生物识别与安全技术研究中心
通讯作者	Wang XB(王晓波)
作者单位	1.京东AI研究院 2.Inception Institute of Artificial Intelligence (IIAI), 3.中国科学院大学 4.中国科学院自动化研究所，模式识别国家重点实验室，生物识别与安全技术中心
推荐引用方式 GB/T 7714	Wu JL,Liao SC,Wang XB,et al. Clustering and Dynamic Sampling Based Unsupervised Domain Adaptation for Person Re-Identification[C]. 见:. 中国-上海. 2019-10.