Abstract: Most of the existing unsupervised domain adaptive pedestrian re-identification methods combine clustering-based pseudo-label prediction with feature fine-tuning.Due to the differences between domains,incorrect pseudo-labels are generated during the clustering process,making pseudo-labels unreliable to a certain extent,misleading feature representation learning,and affec-ting the performance of domain-adaptive models.First,a novel unsupervised domain adaptive network based on counterfactual attention learning is designed,which guides and optimizes the training process by measuring the quality of attention learning,prompting the model to focus on more accurate attention features and reducing the generation of noisy pseudo-labels.Secondly,a noisy samples optimization method based on uncertainty evaluation is proposed.By measuring the inconsistency level between the output features of the student model and the teacher model,as the uncertainty distribution of pedestrian samples in the target domain.The teacher model and the student model are both constructed based on the average teacher method.The uncertainty of the sample is used to reasonably weight each part of the overall loss of the network,and the erroneous influence of the sample with high uncertainty on the overall loss of the model is corrected,and the recognition performance of the target domain is further improved.Experimental data show that the proposed method significantly improves the experimental results in both the source domain DukeMTMC-reID/Market-1501 and the target domain Market-1501/DukeMTMC-reID,with mAP and Rank-1 reaching 82.9%,93.6% and 71.8%,84.4%,respectively.

Key words: Person re-identification, Unsupervised, Domain adaptive, Counterfactual attention, Uncertainty estimation