Towards fast and kernelized orthogonal discriminant analysis on person re-identification. (October 2019)

• Record Type: Journal Article
• Authors: Cao, Min; Chen, Chen; Hu, Xiyuan; Peng, Silong
• Is Part Of: Pattern recognition. Volume 94(2019:Oct.)
• Page Start: 218
• Page End: 229
• Publication Date: 2019-10
• Subjects: Person re-identification -- Metric learning -- Singularity problem -- Orthogonal discriminant analysis; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2019.05.035
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: We propose to solve the singularity problem in person re-identification by learning an orthogonal transformation with the pseudo-inverse of the within-class scatter matrix. We develop a kernel version for learning the orthogonal transformation against the non-linear distribution of data in person re-identification, thereby boosting the performance of person re-identification. We present a fast version with the unchanged performance of person reidentification for improving the solving efficiency. We conduct experiments on four challenging datasets to demonstrates the validity and advantage of the proposed method for solving the singularity problem in person re-identification, and analyze the effectiveness of both kernel version and fast version. Abstract: Recognizing a person across different non-overlapping camera views, is the task of person re-identification. For achieving the task, an effective way is to learn a discriminative metric by minimizing the within-class variance and maximizing the between-class variance simultaneously. However, the dimension of sample feature vector is usually greater than the number of training samples, as a result, the within-class scatter matrix is singular and the metric cannot be learned. In this paper, we propose to solve the singularity problem by employing the pseudo-inverse of the within-class scatter matrix and learning an orthogonal transformation for the metric. The proposed method can be effectively solved with a closed-form solution and no parameters required to tune. In addition, we develop a kernel version against non-linearity in person re-identification, and a fast version for more efficient solution. In experiments, we prove the validity and advantage of the proposed method for solving the singularity problem in person re-identification, and analyze the effectiveness of both kernel version and fast version. Extensively comparative experiments on VIPeR, PRID2011, CUHK01 and CUHK03 person re-identification benchmark datasets, show the state-of-the-art results of the proposed method.