基于多样性约束和离散度分层聚类的无监督视频行人重识别

Unsupervised Video⁃Based Person Re⁃identification Based on Diversity Constraint and Dispersion Hierarchical Clustering

视频行人重识别是一项应用非常广的计算机视觉任务。目前的视频行人重识别方法通常是基于监督学习的,该方法需要手工标记大量的数据,代价非常高且并不适用于现实场景。本文提出了一种从底向上的基于多样性约束和离散度分层聚类的无监督视频行人重识别方法。该方法首先将每个样本当作是一个不同的类,然后结合类内间离散度进行从底向上的分层聚类,类间和类内离散度都小的类别将被优先合并,同时在聚类准则中加入一项多样性约束来平衡每类中的样本数量,最后,利用线性变化的特征存储器动态更新模型。在Mars和DukeMTMC⁃VideoReID两个大型视频数据集上的实验结果表明,相比于目前先进的无监督视频行人重识别方法,本文方法在性能上有一定的提升。

Video-based person re-identification(Re-ID)is a widely used task in computer vision.At present,most video-based Re-ID methods are based on supervised learning,which requires intensive manual annotation,and is very expensive and not suitable for real-life scenarios.In this work,an unsupervised videobased person Re-ID method based on diversity constraint and dispersion hierarchical clustering is proposed.First,each sample is regarded as a single cluster,and both the inter and the intra-cluster dispersions are combined to perform bottom-up hierarchical clustering.Second,clusters with small inter-cluster and intracluster dispersions can be prioritized for merging.At the same time,the diversity constraint is added to the clustering criterion to balance the number of samples in each cluster.Finally,the model is dynamically updated by using a linear feature memory.Experimental results on two public benchmark datasets,including Mars and DukeMTMC-VideoReID,show that compared with the state-of-the-art unsupervised video-based person Re-ID methods,the proposed method has some improvement in performance.