摘要
无监督域自适应行人重识别(Unsupervised Domain Adaptation for person Re-identification, UDA-ReID)任务致力于将知识从已标记的源域数据转移到目标域。和传统的单源域自适应相比,将多源域的知识迁移到目标域是一项更具挑战性的任务。由于领域上的差距,多数据集的简单组合只能产生有限的改进。针对此问题,提出了一种基于精确特征分布匹配和多域信息融合的多源域对比学习(exact feature distribution Matching and multi-domain information Fusion based Multi-domain Contrastive Learning, MFMCL)方法。该方法首先采用具有混合记忆的自步对比学习提取不同域数据的特征,并对提取到的特征进行构图,然后通过两层残差图卷积网络进行多域特征融合。其次,为了增强交叉分布特征、产生更丰富的信息,通过基于排序算法的精确直方图匹配来实现精确特征分布匹配,以获得更多样化的特征增强。实验表明,与目前先进的无监督域自适应行人重识别方法相比,所提出的MFMCL方法在广泛使用的行人重识别数据集Market1501、MSMT17和Duke上都取得了优越的性能。
The task of unsupervised domain adaptation for person re-identification(UDA-ReID)is to transfer knowledge from the marked source domain to the unmarked target domain.Compared with the traditional single-source domain adaptation,it is a more challenging task to transfer multi-source domain knowledge to the target domain.Due to the gap between different domains,the simple combination of multiple datasets can only be improved limited.To solve this problem,a multi-domain contrastive learning method based on exact feature distribution matching and multi-domain information fusion(MFMCL)was proposed.This method firstly extracted the features of data in different domains based on the self-paced contrastive learning with hybrid memory,and constructed the knowledge graph based on the extracted features,and then fused the multi-domain features through the two-layer residual graph convolution network.In addition,in order to enhance the cross distribution features between different domains and generate more abundant image style information,the exact feature distribution matching was realized by accurate histogram matching based on sorting algorithm,so as to obtain more diversified feature enhancement.Compared with the advanced UDA-ReID methods,the experimental results show that our proposed MFMCL method achieves the best performance on the widely used ReID datasets Market1501,MSMT17 and Duke.
作者
杨真真
邵静
杨永鹏
吴心怡
YANG Zhenzhen;SHAO Jing;YANG Yongpeng;WU Xinyi(Key Laboratory of Ministry of Education in Broadband Wireless Communication and Sensor Network Technology,Nanjing University of Posts and Telecommunications,Nanjing,Jiangsu 210003,China;College of Science,Nanjing University of Posts and Telecommunications,Nanjing,Jiangsu 210023,China;School of Network and Communication,Nanjing Vocational College of Information Technology,Nanjing,Jiangsu 210023,China)
出处
《信号处理》
CSCD
北大核心
2023年第6期1099-1107,共9页
Journal of Signal Processing
基金
国家自然科学基金(61501251,62171232)
南京邮电大学宽带无线通信与传感网技术教育部重点实验室开放研究基金(JZNY202113)
南京邮电大学科研项目(NY220207)。
关键词
行人重识别
无监督域自适应
精确特征分布匹配
多域信息融合
图卷积网络
person re-identification
unsupervised domain adaptation
exact feature distribution matching
multi-domain information fusion
graph convolution network
