一种域增强和域自适应的换衣行人重识别范式

Cloth-changing person re-identification paradigm based on domain augmentation and adaptation

为了解决服装变化对行人重识别模型识别人物身份准确率的影响,提出一个基于域增强和域自适应的换衣行人重识别范式,使模型在不同的域中学习通用鲁棒的身份表示特征。首先设计了一种服装语义感知的域数据增强方法,根据人体语义信息,在不改变目标人物身份的情况下,分别改变样本衣服裤子的颜色,生成同人同衣不同色的域数据,填补换衣数据域单一问题;其次设计了一个多正类域自适应损失函数,该函数根据不同域数据在模型训练中所做出贡献的不同,为多正类数据损失赋予不同权重,迫使模型专注于样本的通用身份特征的学习。实验证明,在不影响非换衣行人重识别准确度的情况下,该方法在PRCC和CCVID换衣数据集上的首位命中率和平均精度均值达到了约59.5%、60.0%和88.0%、84.5%。对比于其他方法,这种方法具有更高的准确率和更强的鲁棒性,显著提高了模型识别换衣行人的能力。

In order to solve the influence of the clothing change on the model’s recognition accuracy of the personal identity,a clothes-changing person re-identification paradigm based on domain augmentation and adaptation is proposed,which enables the model to learn general robust identity representation features in different domains.First,a clothing semantic-aware domain data enhancement method is designed based on the semantic information of the human body,which changes the color of sample clothes without changing the identity of the target person to fill the lack of domain diversity in the data;second,a multi-positive class domain adaptive loss function is designed,which assigns differential weights to the multi-positive class data losses according to the different contributions made by different domain data in the model training,forcing the model to focus on the learning of generic identity features of the samples.Experiments demonstrate that the method achieves 59.5%,60.0%,and 88.0%,84.5%of Rank-1 and mAP on two clothing change datasets,PRCC and CCVID,without affecting the accuracy of non-clothing person re-identification.Compared with other methods,this method has a higher accuracy and stronger robustness and significantly improves the model’s ability to recognize persons.