Contrastive Learning Based Unsupervised Person Re-Idendification

In recent years,due to the application requirements of smart cities and security,person re-identification has received extensive attention in the field of practical application and academic research of computer vision.The goal of person re-identification is to retrieve images or video clips that match the images to be retrieved from many cross-camera images and video streams.However,recent research in person re-ID has focused on the field of supervised learning,which relies heavily on person labeling.Therefore,based on the perspective of unsupervised learning,this paper proposes an efficient and high-precision person re-identification model and training method.With the emergence of parameter-free instance-level classification methods and a large number of contrastive learning methods,this paper proposes the following innovative works based on these methods:First,this thesis proposes an unsupervised training framework of Hybrid Contrastive Learning(HCL),which combines the advantages of instance-level contrastive learning and cluster-level contrastive learning,and not only ensures sufficient learning of the relationship between positive sample pairs,which provides sufficient negative samples to further optimize the entire contrastive learning framework.In addition,based on the HCL framework,this paper proposes a Multi-Granularity Cluster Ensemble(MGCE)method,which further improves the performance of the HCL framework by utilizing clustering information of different granularities.The paper conducts sufficient experiments on the person reidentification benchmark datasets Market-1501 and DukeMTMC-reID,and the experimental results also fully verify the effectiveness of the MGCE-HCL framework.Second,based on MGCE-HCL,the thesis proposes Multi-level Contrastive Learning(MCL)and Adaptive Multi-Granularity Cluster Ensemble(adaMGCE)methods to improve the effect of the above methods.The method is also tested on Market-1501 and DukeMTMC-reID,and the experimental results demonstrate the effectiveness of the method.