Unsupervised Cloth-changing Person Re-identification Algorithm Based on Multi-grained Feature Network
Aiming at the problems of data labeling difficulties and clothing changes in person re-identifica-tion tasks,a multi-grained feature network(MGFNet)based on unsupervised method PPLR(Part-based Pseudo Label Refinement)was proposed.Only the RGB images were input to fully extract global,local and facial features of persons in the images,and the clothing information in the features was suppressed according to the attention mechanism to mine the substantive features of persons.The fusion of global,fa-cial and local features was followed by the clustering algorithm to generate precise pseudo-labels,which were used to supervise model training.The performance of MGFNet was evaluated on public cloth-chan-ging datasets,and ablation studies were designed to validate the effectiveness of MGFNet.The results in-dicate that MGFNet's mAP and Rank-1 metrics on the PRCC dataset respectively improved by 6.9％and 15.8％compared to the baseline model PPLR.
cloth-changing person re-identificationunsupervised learningdual-branch networkattention mechanismfeature alignment
万方数据
