Large Scale Person Re-Identification

Person Re-Identification(Re-ID)aims to solve the problem of detecting specific persons across different cameras and scenes using computer vision techniques.This task has important research implications and a wide range of applications,such as for abnormal behavior detection,assisted judicial forensics,and cross-camera tracking.It is one of the hot topics in the field of computer vision,which is attracting more and more researchers' interests.Person Re-Identification has typical challenges from the very beginning of its def-inition,such as low resolution of images,variations in viewpoint and lighting,different pedestrian postures,and variations in clothing,etc..The emergence of deep learning has largely alleviated these typical problems.However,deep learning-based person re-identification has its own limitations,such as the demand for large amounts of labeled training data which is difficult to obtain in real-world scenarios.Toward solving the problems in large-scale person re-identification,this paper makes a large variety of ex-plorations.From the perspective of training,we propose novel large-scale unsupervised and weakly supervised pre-training methods.Meanwhile from the perspective of large-scale application for person re-identification,we propose an efficient post-processing algorithm.The main contributions are summarised as follows:1.For the lack of large-scale training data in person re-identification,the first large scale person re-identification dataset,LUPerson,is proposed,which reaches the scale of ImageNet.To avoid the high cost of manual labeling,LUPerson is constructed in an unsupervised manner,which also allows this dataset to be easily scaled up.Based on LUPerson,this paper is the first to successfully accomplish large-scale unsuper-vised pre-training for person re-identification successful,and explores several key fac-tors that make unsupervised person re-identification pre-training successful.Finally,a pre-trained model with extremely strong representative ability is obtained,which can significantly improve the performance of existing benchmarks by just making simple substitutions without additional cost.2.Based on LUPerson,this paper uses the state-of-the-art tracking algorithm to explore the time-space information in the original video to obtain the person tracklets,and relies on the tracklets to assign weak labels to every pedestrian,and then constructs a large-scale weakly supervised person re-identification dataset,named LUPerson-WS.As the annotations in LUPerson-WS are not consistent and clean,a new weakly super-vised pre-training algorithm,WSP,is proposed to address the problem that the anno-tations in LUPerson-WS are very noisy,and a stronger pre-training model is obtained based on WSP trained on LUPerson-WS,which can push the performance of existing benchmarks to a new limit.Also,the weakly supervised pre-training model is extremely superior when only a small amount of training data is available.3.Post-processing algorithms can significantly improve the performance of the original model at the time of testing with a small cost.This paper proposes a post-processing method based on impression updating that can iteratively adjust the per-ception of a person and thus improve the accuracy of retrieval.The post-processing algorithm is simple in design and has fast speed while taking into account performance,and can excellently meet the needs in large-scale practical application scenarios.In conclusion,this paper presents novel solutions and pioneering methods for the training and post-processing of large-scale person re-identification.Both the pre-training and post-processing methods proposed in this paper are very flexible and gen-eral,which can be widely applied to various state-of-the-art person re-identification methods and significantly improve the performance of these algorithms.We believe that the proposed methods will boost the practical potential of person re-identification in real-world scenarios.