Unveiling the Potential of Vision Transformer Architecture for Person Re-identification

Abstract

Person re-identification (Re-ID) is a process to re-identify a person if he has been already seen by a camera network. Since start the convolutional neural networks (CNNs) are dominantly being used to solve the person Re-ID problem. The default limitation of CNNs i.e., local receptive field, prohibits the network to learn the distinctive global dependencies at initial layers. This study proposes a self-attention based deep architecture that learns global dependencies at each network layer to address CNN's limitation. Additionally, the introduction of a novel contextual learning module called Attention Drop Block (ADB) supports learning of less attentive areas of an image as well. The proposed model is evaluated on two public Re-ID benchmarks Market1501 and DukeMTMC-ReID, and outperformed all CNN baseline Re-ID models. The implementation and trained models are made publicly available at https://git.io/JYRE3.