Joint learning with diverse knowledge for re-identification

Abstract

Re-identification (re-ID) aims to search the target images of pedestrians or vehicles with the same identity in the non-overlapping camera network. Some factors, such as, cluttered background, different illumination and occlusion, which increases the difficulty of the re-identification task. To fulfill the re-ID task, this paper presents a re-ID method with diverse knowledge, which contains a multi-branch structure to extract various detailed features of samples from multiple perspectives. Furthermore, the attention mechanism is introduced to suppress the influence of background information. Moreover, to improve the discriminative ability, a joint learning strategy is presented to refine the network and align the output of multiple branches. To verify the effectiveness of the proposed method, corresponding experiments have been conducted on multiple public datasets and experimental results demonstrate that our method is effective and achieves the competitive results with existing methods on CUHK03 and Market1501.