基于反事实注意力学习的无监督域自适应行人再识别

Unsupervised Domain Adaptive Pedestrian Re-identification Based on Counterfactual AttentionLearning

现有的无监督域自适应行人重新识别方法大多是将基于聚类的伪标签预测与特征微调相结合。由于域间存在差异,聚类过程中产生了不正确的伪标签,使伪标签存在一定的不可靠性,误导特征表示学习,从而影响域自适应模型的性能。基于此,首先设计一个新颖的基于反事实注意力学习的无监督域自适应网络,通过衡量注意力学习的质量对训练过程进行指导优化,促使模型关注更加精准的注意力特征,减少噪声伪标签的生成;其次提出了一种基于不确定性评估的噪声样本优化方法,通过测量基于平均教师方法的学生模型和教师模型输出特征之间的不一致性水平,将其作为目标域行人样本的不确定性分布,进而利用样本的不确定性对网络总体损失的各个部分进行合理加权,修正具有高不确定性的样本对模型总体损失的错误影响,进一步提升目标域的识别性能。实验数据表明,所提方法在源域DukeMTMC-reID/Market-1501和目标域Market-1501/DukeMTMC-reID上的实验结果都有显著提高,mAP和Rank-1分别达到了82.9%,93.6%和71.8%,84.4%。

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.