Cross-view semantic projection learning for person re-identification. (March 2018)

• Record Type: Journal Article
• Authors: Dai, Ju; Zhang, Ying; Lu, Huchuan; Wang, Hongyu
• Is Part Of: Pattern recognition. Volume 75(2018:Mar.)
• Page Start: 63
• Page End: 76
• Publication Date: 2018-03
• Subjects: Person re-identification -- Feature transformation -- Semantic representation learning -- Semantic projection learning; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2017.04.022
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: We propose a novel feature transformation algorithm for person re-identification. The hand-crafted features are mapped into the common semantic space via view-specific semantic projection functions. The common intrinsic structure is explored via a shared semantic basis matrix. The association matrix is utilized to capture the optimal associations of the same persons. We further extend the CSPL algorithm for multi-view analysis. Abstract: Feature transformation is of great importance to strengthen the descriptive power of feature representation for many classification and recognition tasks. In this paper, we propose a novel cross-view semantic projection learning method for extracting latent semantics from the hand-crafted features. Specifically, the shared latent basis matrix, the view-specific semantic projection functions and the optimal associations of different views are jointly learned in a unified matrix factorization framework, to get a common semantic space where images of the same person can be well characterized. We further present a generalization of the approach to multiple views. Extensive experiments on a series of challenging datasets highlight the superiorities of the proposed algorithm and demonstrate the effectiveness of the generalized version in multi-view person re-identification applications.