Unsupervised Person Re-Identification Based On Camera Information And Mutual Information Maximization

Person re-identification is an important technology in intelligent surveillance video analysis.Given a query person image,the purpose of person re-identification is to retrieve images or frames containing the same person from the image or video database.At present,the research of person re-identification based on supervised learning has made significant progress.However,when this kind of methods are used to test on another dataset,the results obtained are not ideal.In addition,the annotation of person reidentification datasets is very time-consuming and laborious.In order to improve the scalability of the person re-identification algorithms and get rid of the dependence on manual annotation,unsupervised person re-identification has become a research hotspot at home and abroad in recent years.This paper conducts a more in-depth study on this issue.We first propose an unsupervised person re-identification algorithm based on camera information.This method is based on two observations.Firstly,most of current unsupervised person re-identification algorithms are based on domain adaptation,relying on labeled source dataset and the similarity between the source dataset and the target dataset.Secondly,camera labels can be automatically obtained during the collection process of datasets.They provide additional information for the video.This method aims to use camera labels to make the model more discriminative under the setting of fully unsupervised learning,so as to get rid of the dependence on auxiliary dataset.This paper proposes a new learning strategy that divides the originally difficult learning process into three relatively easy stages.It performs unsupervised person re-identification through initialization with instance discrimination,intra-camera learning and inter-camera learning.The three stages share the parameters of the backbone network with each other.We conducted experiments on the commonly used datasets for person reidentification,and the experimental results show that this method surpasses other methods under the same problem setting.In addition,this paper also studies the application of deep mutual information maximization in unsupervised person re-identification algorithms.The method proposed in this paper estimates and maximizes the mutual information between the person image and the person feature representation on the basis of the baseline model,so that the person feature can retain the ID information contained in the original image as much as possible,thereby improving the model’s performance.Experimental results verify the effectiveness of the method.This method provides a new angle for the application of deep mutual information maximization and contrastive learning in unsupervised person re-identification algorithms.