基于异构卷积神经网络集成的无监督行人重识别方法

An Unsupervised Person Re-Identification Method Based on Heterogeneous Convolutional Neural Networks Ensemble

行人重识别旨在从不同的摄像头中识别目标行人的图像.由于不同场景之间存在域偏差,在一个场景中训练好的重识别模型无法直接应用在另一个场景中.为克服该问题,现有的无监督行人重识别方法倾向通过使用聚类算法获得伪标签,再利用伪标签训练重识别模型.但是,由于聚类结果是不准确的,这类方法会引入大量噪声标签,从而限制了模型的泛化能力.因此,为减轻噪声伪标签的影响,本文提出了一种基于异构卷积神经网络集成的无监督行人重识别方法.该框架不使用任何人工标记信息,自动推测目标域中行人图像之间的关系,并构建协作可信实例选择机制,选择可信度高的伪标签用于模型的训练.通过设计双分支异构卷积神经网络学习判别能力强的多种行人特征,并利用记忆单元存储训练过程中的全局特征,减少因噪声标签在训练过程中产生的波动,提高模型的鲁棒性.本文方法在多个公开行人数据集上进行了验证并得到了良好的实验结果.在Market1501和DukeMTMC-reID数据集上,mAP分别达到了85.4%和74.8%.

Person re-identification(re-ID)aims to identify a person's images across different cameras.However,the domain bias between different datasets makes it a challenge for re-ID models trained on one dataset to be adapted to anoth⁃er.A variety of unsupervised domain adaptation methods tend to transfer learned knowledge from one domain to another by optimizing with pseudo-labels.However,these methods introduce a large number of noisy labels through one-shot clus⁃tering,which hinders the retraining process and limits generalization.To mitigate the impact of noisy pseudo-labels,this pa⁃per proposes an unsupervised person re-identification method based on an ensemble of heterogeneous convolutional neural networks.The framework does not apply any manual labeling information,automatically infers the relationship between pe⁃destrian images in the target domain,and a cooperative trusted instance selection mechanism is established to select pseudolabels with high credibility.By constructing a dual-branch heterogeneous network,a variety of different pedestrian features are learned,and memory structures are designed to store the life-long features during the training stage,which could reduce the fluctuation of noise labels,and improve the robustness of the model.Comprehensive experimental results have demon⁃strated that our proposed method can achieve excellent performances on benchmark datasets.And mAP is increased to 85.4%and 74.8%on Market1501 and DukeMTMC-reID,respectively.