Two-stage person re-identification scheme using cross-input neighborhood differences

Person re-identification aims to identify images of a particular person captured from different cameras or the same camera under different conditions. Person re-identification is conducted using an identification model that classifies the identity of the selected person or a verification model that discriminates between positive and negative image pairs. To further improve the re-identification performance, various methods have combined identification loss with verification loss. However, because such methods compare identities using one-dimensional embedding features without spatial information, local relationships are not considered. Thus, in this paper, we propose a two-stage person re-identification scheme using feature extraction and feature comparison networks. The former generates feature maps with spatial information, and the latter calculates their neighborhood and global differences. We conducted extensive experiments using well-known person re-identification datasets, and the proposed model achieved rank-1 accuracies of 84% and 88.4% for CUHK03 and Market-1501, respectively.