A discriminative null space based deep learning approach for person re-identification

Person re-identification across multiple camera views is a rather challenging task due to various view points, illuminations, backgrounds and poses. How to extract discriminative features is the most critical way to overcome these challenges. In this paper, we design a discriminative null space based deep learning approach for person re-identification. Firstly, a Siamese Convolutional Neural Network (SCNN) is designed to automatically learn effective semantic features for person re-identification in different camera views. Furthermore, to obtain better recognition performance, we adopt the Null Foley-Sammon Transform (NFST) metric learning approach to combine the low-level, mid-level features and high-level features learned by the SCNN in a new discriminative null space. In this null space, images of the same person are collapsed into a single point thus minimizing the within-class scatter to the extreme and maximizing the relative between-class separation simultaneously. Finally, the comprehensive evaluations demonstrate that our approach outperforms all state-of-the-art methods on the Market-1501, which is the world's largest person re-identification benchmark dataset.