Efficient Feature Learning For Unsupervised Person Re-identification

In recent years,it has led to an increased emphasis on unsupervised person re-identification techniques that eliminate the need for manually labeled data.However,the fine granularity of the person image and the variations in illumination,background,and camera angle prevent deep neural networks from learning effective pedestrian features in unsupervised settings.To solve the unsupervised person re-identification problem,it is crucial to distinguish between pedestrians so that a deep neural network can accurately learn useful features and solve the problem of unsupervised person re-identification.In order to learn the effective features of pedestrian images,this work analyzes the main limitations of current unsupervised person re-id techniques and focuses on improving clustering techniques,enhancing training strategies,embedding auxiliary information,and building network frameworks.The principal innovations and contributions of this paper are listed below:(1)This study proposes SPCNet-SI,which embeds auxiliary information into unsupervised model training.It employs camera IDs as auxiliary information to assist the network in learning features across camera views and incorporates self-paced learning into the network training process to reduce the hindrance of camera views to effective feature learning.SPCNet-SI performs admirably on multiple person re-ID datasets.(2)This study proposes C3AB,which combines clustering consolidation and cluster adaptive balancing loss to effectively train the network by screening noisy samples within clusters and adaptively assigning sample weights.Experiments on multiple person re-id datasets,including Market-1501,demonstrate that the existing baseline model combined with the C3AB method can significantly improve performance,achieving the highest level of unsupervised person re-id methods at the time.(3)This study proposes a cluster-guided asymmetric contrastive learning framework called CACL,which aims to extract consistent features from multiview samples.CACL employs asymmetric data augmentation for contrastive learning to encourage the model to learn more discriminative fine-grained features while suppressing the influence of image color dominance.Experimental results on several datasets,including MSMT17,demonstrate that the CACL significantly enhances the ability to distinguish pedestrians wearing similarly colored clothing.(4)This study proposes a semantic mask-driven contrast learning method,called MasKCL.MasKCL embeds the person silhouette mask as a semantic prompt into contrastive learning and combines semantic information with RGB information to construct a hierarchical neighbor structure that drives the model to learn clothing-independent effective features.Experiments demonstrate that the performance of the proposed model is significantly higher than that of unsupervised short-term person re-id methods and close to that of supervised longterm person re-id methods,and resolves the problem of retrieving pedestrians with changing clothes in unsupervised settings.