基于自适应融合网络的跨域行人重识别方法

Cross-domain Person Re-identification on Adaptive Fusion Network

无监督跨域的行人重识别旨在将从有标签的源域中学习到的知识迁移到无标签的目标域,具有实用性和有效性而得到广泛关注.基于聚类的跨域行人重识别可以生成伪标签并对模型进行优化使得其表现较其他方法更优,然而这类方法由于过于依赖聚类伪标签的准确性,忽略了对伪标签噪声的处理,导致噪声随着网络迭代而不断扩大,影响模型的鲁棒性.针对这个问题,提出了基于自适应融合网络的方法,利用双网络结构共同学习,并将学习到的知识进行融合得到融合网络;为了区分两个网络的学习能力,设计了自适应融合策略;同时,利用细粒度风格转换模块对目标域数据集进行处理,降低行人图像对相机变换的敏感度.在行人重识别基准数据集Market1501、DukeMTMC-ReID和MSMT17上,通过评估指标平均精度均值和Rank-n与主流的方法进行了对比实验,验证了该方法的有效性.

Unsupervised cross-domain person re-identification aims to transfer the knowledge learned from labeled source domain to unlabeled target domain,which has attracted wide attention due to its practicability and effectiveness.Cross-domain person re-identification based on clustering can generate pseudo-labels and optimize the model to make its performance better than other methods.However,these methods rely too much on the accuracy of clustering pseudo labels and ignore to deal with pseudo-label noise,which leads to the continuous expansion of noise with network iteration and affects the robustness of the models.To address this problem,this paper proposes a method based on fine-grained style transfer and adaptive fusion network,which uses dual network structure to learn together and fuse the learned knowledge to obtain a fusion network.To treat the learning ability of the two networks differently,an adaptive fusion strategy is designed based on the different weights of the two networks in each fusion process.At the same time,a fine-grained style transfer module is used to process the target domain dataset,thereby reducing the sensitivity of person images to camera transformation.On the person re-identification benchmark datasets Market1501,DukeMTMC-ReID and MSMT17,the effectiveness of the proposed method was verified by comparing mean average precision and Rank-n with the state-of-the-art methods.