Structure alignment of attributes and visual features for cross-dataset person re-identification. (October 2020)

• Record Type: Journal Article
• Authors: Li, Huafeng; Kuang, Zhenyu; Yu, Zhengtao; Luo, Jiebo
• Is Part Of: Pattern recognition. Volume 106(2020:Oct.)
• Publication Date: 2020-10
• Subjects: Person re-identification -- Self-supervised strategy -- Domain adaptation -- Structure alignment -- Self-reconstruction; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2020.107414
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: Self-reconstruction structure alignment is proposed for cross-domain person re-identification. Visual attribute are aligned by class prototype to promote discrimination of predicted attributes. A self-supervised learning framework is developed to alleviate the domain bias. Abstract: In cross-dataset person re-identification, it is challenging to address the problem of domain shift between training and test data. Although unsupervised domain adaptation methods have been developed, the performance is still much weaker compared with that of supervised methods because these models cannot follow a supervised optimization in unlabeled target domains. To address this problem, a transductive structure alignment-based self-reconstruction dictionary learning approach is proposed in this paper for cross-dataset person re-identification (PRID). Specifically, visual-attribute embedding is first learned to achieve knowledge transfer from the source domain to the target domain. In this process, visual-attribute structures are aligned via class prototype dictionaries to promote the discrimination of predicted semantic attributes by exploiting structure information between the visual feature and class prototype. Moreover, to mitigate domain shift, domain-invariant visual-attribute self-reconstruction is integrated into our dictionary learning framework. An identifier is then constructed by integrating the discriminativeness of attribute and compatibility matrix shared both source domain and target domain. Finally, the pre-learned model is tuned by selecting samples from the target domain which are not labeled but assigned pseudo-labels. Extensive experimental results on benchmark datasets show that our approach outperforms several state-of-the-art approaches.