CD-ABM: Curriculum Design with Attention Branch Model for Person Re-identification

Abstract

Person re-identification (re-ID) is a challenging problem due to background clutter, illumination and pose variation, occlusion, and pedestrian misalignment. Current state-of-the-art methods commonly extract discriminative information by deep networks based on one-stage training. Though straightforward, using one-stage learning, the presence of pedestrian misalignment in practical applications may significantly degrade the performance of the learned model. To address this issue, we propose a novel model for person re-ID, called CD-ABM. It adopts a curriculum design to proceed training from easy to hard samples and generates an attention map in a supervised manner to further facilitate discriminative feature extraction. Compared with existing methods, CD-ABM has the following advantages: (1) The curriculum design can gradually improve the model capability through progressive learning. (2) The attention map enables the local branch to be associated with the global branch and better exploits both local and global information. Experiments on three benchmark datasets show that, CD-ABM can achieve competitive performance with the state-of-the-arts. Noteworthily, on the most challenging dataset MSMT17, it surpasses state-of-the-art methods by 15.9% in Rank-1 and 21.0% in mAP.