Person Re-identification Via Dual-path Attention

Aiming at the problem that the existing methods for person Re-ID pay insufficient attention to the non-salient distinguishable features and can’t extract the strong expression ability of pedestrian features. a Re-ID method (DPANet) based on feature extraction of dual-path attention mechanism is proposed. DPANet consists of a dual-path attention backbone network and an enhanced attention feature fusion module. Among them, the dual-path attention network enables the model to focus on feature regions with different levels of saliency, to mine significant and potentially non-significant discriminable features respectively, and to emphasize the importance of potential key features. The attentional feature fusion module is enhanced to complete feature information complementation, while counterfactual intervention is used to strengthen the quality and validity of the acquired attentional feature maps, so as to obtain a more discriminative final feature representation. The experimental results show that, on the Market1501, DukeMTMC-reID and MSMT17 datasets, the mAP values reached 89.3%, 80.0%, 58.4% and the Rank-1 values reached 95.7%, 89.8%, 80.7%, respectively, fully proving the effectiveness of this method.