Contrastive Learning Based Person Re--identification

Person re-identification(re-ID)is the task of matching the same person across non-overlapping camera views.Given a query person image,person re-ID aims to find in other camera views the images where the person appears.In recent years,owing to the wide application of person re-ID in smart video surveillance,criminal investigation and public security system,researchers have conducted extensive research on person re-ID.By exploiting the person annotation infor-mation,fully-supervised person re-ID has achieved quite high accuracy and has been widely used in real-world applications.However,fully-supervised person re-ID requires the pairwise comparison of images in order to obtain the person annotation.When the dataset scale becomes larger,the heavy annotation consumes an overwhelming amount of labor resources.In order to reduce the annotation cost and facilitate the fast application of person re-ID models,it is of high importance to conduct research on person re-ID under limited annotation scenarios.As a task of retrieval and matching,the characteristic of person re-ID is that the train-ing and test dataset contain different set of persons.By performing model learning with the train set,the main goal is to learn discriminative features that can well distinguish different persons.Further,it is expected that the learned feature can generalize on unseen test dataset and separate unseen persons well.In recent years,contrastive learning has shown impressive performance on unsupervised and supervised learning tasks.By comparing an instance with its positive and negative comparisons,contrastive learning tries to increase the similarity between an instance and its positive comparisons,while decrease the similarity between an instance and its negative comparisons.As a result,contrastive learning can help model learn the intra-class similarity and inter-class difference in both supervised and unsupervised settings.More specif-ically,performing similarity learning among different instances or classes enables the model to learn the high-level semantic relations in input data,thus generating features that improve intra-class compactness and inter-class dispersion.The property of contrastive learning actu-ally quite matches the goal of person re-ID task,which is to learn highly discriminative features that generalize to unseen data.Therefore,inspired by the contrastive learning perspective,this thesis proposes effective methods under intra-camera supervised as well as unsupervised re-ID scenarios,in order to tackle the re-ID task with limited annotation.Specifically,the main contributions of this thesis are as follow:· This thesis performs a preliminary study on the characteristic of contrastive learning for person re-identification task.Under different re-ID scenarios,this thesis compares para-metric classification loss and triplet loss with non-parametric classification loss in con-trastive learning.By comparison and analysis,the characteristic and variation of different losses are derived.Extensive experiments are conducted to support the analysis.· This thesis proposes a graph-induced contrastive learning method for intra-camera super-vised(ICS)person re-ID.Based on the specific association constraints under ICS re-ID scenario,this method proposes a graph partitioning algorithm to predict reliable person associations.Based on the association result,the method further proposes a progressive contrastive loss,which performs intra- and inter-camera contrastive learning in a uni-fied learning framework,and effectively improves the discrimination ability of the re-ID model.· This thesis proposes camera-aware proxies for unsupervised person re-ID task.By con-sidering the intra-cluster distribution variance caused by camera discrepancy,this method proposes to learn camera-aware proxies to capture the local structure insider each cluster.Based on the camera-aware proxies,the method further proposes complementary intra-camera and inter-camera contrastive losses,in order to optimize the relations between instances and the proxies in a contrastive learning manner,so that the model discrimi-nation is enhanced.Experimental results on multiple benchmark datasets show that,the proposed method achieves higher performance compared to the state-of-the-art methods,and greatly improves the accuracy for unsupervised person re-ID task.· This thesis proposes an unsupervised re-ID method based on proxy association enhance-ment.This method focus on the existing problems in the former camera-aware proxies method,which is the over-clustering phenomenon,and designs an enhancement model for optimization purpose.To make up for the association inefficiency of unsupervised clustering,and to enhance the cross-camera associations,this method designs a similarity-based per-camera hard positive association mining strategy.The proposed strategy pre-dicts effective cross-camera positive associations for supervising the model.Experimen-tal results on several re-ID benchmarks demonstrates the effectiveness of the proposed method.