基于多对多生成对抗网络的非对称跨域迁移行人再识别

Asymmetric Cross-domain Transfer Learning of Person Re-identification Based on the Many-to-many Generative Adversarial Network

无监督跨域迁移学习是行人再识别中一个非常重要的任务.给定一个有标注的源域和一个没有标注的目标域,无监督跨域迁移的关键点在于尽可能地把源域的知识迁移到目标域.然而,目前的跨域迁移方法忽略了域内各视角分布的差异性,导致迁移效果不好.针对这个缺陷,本文提出了一个基于多视角的非对称跨域迁移学习的新问题.为了实现这种非对称跨域迁移,提出了一种基于多对多生成对抗网络(Many-to-many generative adversarial network,M2M-GAN)的迁移方法.该方法嵌入了指定的源域视角标记和目标域视角标记作为引导信息,并增加了视角分类器用于鉴别不同的视角分布,从而使模型能自动针对不同的源域视角和目标域视角组合采取不同的迁移方式.在行人再识别基准数据集Market1501、DukeMTMC-reID和MSMT17上,实验验证了本文的方法能有效提升迁移效果,达到更高的无监督跨域行人再识别准确率.

Unsupervised cross-domain transfer learning is an extremely important task in person re-identification(ReID).Given a labeled source domain and an unlabeled target domain,the key to the unsupervised cross-domain transfer learning is to transfer the knowledge from the source domain to the target domain as much as possible.However,current cross-domain transfer learning methods cannot obtain desired performance because they ignore the distribution differences between different views within domains.Therefore,we propose a new problem of view-based asymmetric cross-domain transfer learning for ReID.To address this problem,we propose a novel transfer learning method based on the many-to-many generative adversarial network(M2M-GAN).The M2M-GAN embeds source view labels and target view labels as the guide information,and adds view classifiers to identify different view distributions,so that the model can automatically adopt different transferring ways according to different source views or target views.Experiments on three ReID benchmark datasets Market1501,DukeMTMC-reID and MSMT17 verify that the proposed method can improve the performance of transfer learning and achieve higher recognition rate of unsupervised cross-domain ReID.