Unsupervised Person Re-Identification Based On Reliable Pseudo Labels

Video surveillance systems have come into our daily life and play an important role in the public security field.Currently,we mainly rely on the human to filter out specific person from huge amount of vides,which is not only inefficient and unable to work in real time,but also a waste of time.Person re-identification has attracted a lot of interest which aims to match targets across cameras,and it can quickly retrieve targets in multi cameras.In recent years,benefiting from the development of deep learning,supervised person re-identification methods based on neural network have made significant progress and the performance has surpassed that of the professionals.However,the supervised methods require a large amount of labeled data,which increases the cost.To address this problem,unsupervised methods get rid of the dependence on annotation information by predicting pseudo labels or using domain adaptation methods,but there is still a big gap compared with supervised methods.Most studies focus on improving the network’s ability to deal with noisy labels or extracting more discriminative features from the unlabeled datasets,while ignoring the reliability of the pseudo labels generated by the clustering algorithm itself.This paper focuses on how to generate reliable pseudo labels for unsupervised person re-identification.This work takes samples as nodes of undirected graph and try to enhance the reliability of pseudo labels with the help of maximum clique algorithm as a strict constraint.Another work is to divide the data into several domains according to the camera index of the samples,predict the pseudo labels separately and then merge between the domains.The main contributions of this work are as follows:1.This paper proposes a pseudo label generation method based on maximum clique algorithm.The method takes the samples in the target datasets as nodes of the graph,then find out the maximum cliques on each connected component,finally merges the cliques from different domain,and finally calculate the probability of each class belonging to the remaining cliques and assign them pseudo labels.The experiments indicate that it is feasible to generate pseudo labels based on the maximum clique algorithm.2.The unlabeled dataset is divided into multiple camera domains,and the maximum clique algorithm is used to generate pseudo labels in the domain,and then the pseudo labels between domains are merged.The influence of inter-camera variance can be avoided by dividing the dataset into serval camera domains when performing maximum clique algorithm,and the first match rate on the public dataset reaches 89%,exceeding the best method by 2.8%.3.An application is used to test the proposed method in wild world scenarios.A person re-identification model is constructed with the method proposed in this paper,the data sources are 16 cameras in public,and the efficiency and accuracy of the algorithm in retrieving target pedestrians in surveillance videos are tested.The results show that the algorithm proposed in this paper can be used in real scenes.