Semi-Coupled Dictionary Learning With Relaxation Label Space Transformation for Video-Based Person Re-Identification

Abstract

Video-based person reidentification (re-id) is a challenging problem due to much discrepancy between different videos by person pose, illumination, viewpoint change, background clutter, and occlusion within each camera and across different cameras. However, most existing video-based person re-id methods usually focus on dealing with the discrepancy between different cameras and do not fully consider the correlation between different cameras. In this paper, we propose a semicoupled dictionary learning with relaxation label space transformation approach to capture the intrinsic relationship of the same person under different cameras. First, to reduce the discrepancy between different views, we transform the original feature spaces into the common feature space by local Fisher discriminant analysis. Two dictionaries are learned from this common feature space. Second, a relaxation label space is introduced to associate the same person under different views. In this label space, the distance between different persons can be enlarged as much as possible, such that label information has stronger discriminative capability. A single dictionary is learned from the relaxation label space. Finally, in order to further enhance the correlation of the same person between different cameras, we use a pair of transformation matrices which map the coding coefficients learned from the common feature space to the coding coefficients learned from the relaxation label space, respectively. Extensive experimental results on two public iLIDS Video re-IDentification and Person Re-ID 2011 video-based person re-id datasets demonstrate the effectiveness of the proposed method.