Abstract
The goal of occluded person re-identification (ReID) is to retrieve specificpedestrians in occluded situations. However, occluded person ReID still suffersfrom background clutter and low-quality local feature representations, whichlimits model performance. In our research, we introduce a new framework calledPAB-ReID, which is a novel ReID model incorporating part-attention mechanismsto tackle the aforementioned issues effectively. Firstly, we introduce thehuman parsing label to guide the generation of more accurate human partattention maps. In addition, we propose a fine-grained feature focuser forgenerating fine-grained human local feature representations while suppressingbackground interference. Moreover, We also design a part triplet loss tosupervise the learning of human local features, which optimizesintra/inter-class distance. We conducted extensive experiments on specializedocclusion and regular ReID datasets, showcasing that our approach outperformsthe existing state-of-the-art methods.