Attention-based mechanism and feature fusion network for person re-identification

For the problem that person features cannot be sufficiently extracted in person re-identification, a person re-identification model based on attention mechanism is proposed. Firstly, person features are extracted using a hybrid network combining transformer's core multi-headed self-attentive module with the convolutional neural network ResNet50-IBN-a. Secondly, an efficient channel attention mechanism ECANet is embedded to make the model of this paper more focused on the key information in the person foreground. Finally, fusing the mid-level and high-level features in the model can avoid some discriminative features loss. The experimental results show that the provide model achieves 94.8% rank-1 and 84.5% mAP on the Market-1501 dataset; achieves 84.9% rank-1 and 65.9% mAP on the DukeMTMC-reID dataset; and achieves 40.3% rank-1 and 33.3% mAP on the Occluded-Duke MTMC dataset. Our proposed model compares well with some of the existing person re-identification models on these datasets mentioned above.