Research On Unsupervised Person Re-identification Method Based On Contrastive Learnin

Person re-identification is a crucial technique for realizing a smart surveillance system,which aims to identify the same person from person images taken by different cameras and has great application value in the field of public security.According to the differences in data settings,the extant person re-identification methods can be divided into supervised methods and unsupervised methods.As unsupervised person re-identification methods do not require additional time and labor to annotate datasets,they are more practical and popular than supervised person re-identification approaches.Currently,the state-of-art unsupervised person re-recognition methods mainly use clustered pseudo labels to construct proxies for contrast learning so as to obtain discriminative person matching models.In order to further improve the accuracy of person image matching under the unsupervised setting,this paper conducts the following researches against the drawbacks of the existing contrast learning methods:(1)Existing contrast learning methods only utilize local structures within IDs to design their proxies while ignoring the relations between samples of different IDs,which limits the improvement for inter-ID discriminative ability.To resolve this issue,this paper proposes a Global Relation-Aware Contrast Learning method for the task of unsupervised Re-ID.This method first sets up two proxies for each cluster to capture the inter-and intra-ID relations respectively,which enables us to both effectively increase inter-ID variances and reduce the intra-ID discrepancies.Specifically,the samples that are most different from those in different clusters are selected as inter-ID relation-aware proxies,while those that are least similar to samples from the same clusters are employed as intra-ID relation-aware proxies.With the aid of these proxies,this method designs both inter-and intra-ID relation-aware contrastive learning modules to facilitate feature learning.By pulling each sample close to the positive proxy,this method can obtain identity-invariant discriminative features.Finally,comparison experiments on five widely-used Re-ID datasets prove that this method can achieve performance advantages.Specifically,on the hardest MSMT17 dataset,this method can obtain7.6% Rank-1 improvements and 7.7% mAP improvements.(2)Previous contrast learning methods simultaneously address intra-ID discrepancies under all cameras and require independent learning under each camera,which increases the complexity of algorithm.To resolve this issue,the paper present a camera contrast learning framework for unsupervised person Re-ID.This method first proposes a time-based camera contrastive learning module to facilitate model learning.At each iteration,this module follows the time contrast principle to select one camera centroid as proxy of each cluster.By enforcing the samples to converge to positive proxies,the correlation between features and cameras can gradually be reduced.Moreover,this method designs a 3-dimensional attention module to further reduce intra-ID discrepancies caused by background shifts under different cameras.By re-weighting each feature map element in a spatial-channel order,this module can exactly find identity-invariant semantic cues from regions of interest in person images,no matter how the background change.Experimental results on five popular datasets demonstrate the superiority of this method for the unsupervised person re-identification task.Notably,on the largest MSMT17 dataset,this method can improve the Rank-1 and mAP accuracy by 4.0% and 4.9%,respectively.