Person Re-identification Based On Style Transfer

Nowadays,surveillance cameras are distributed in every corner of life.To solve person recognition and retrieval under cross-camera,person re-identification technology has become a hot topic in the field of computer vision,and it is of great significance in intelligent security,album clustering,and intelligent business.However,there are variations in the frame,position,and illumination between different camera devices and the appearance of the same pedestrian in different clothes and pose will also be different.To overcome the above-mentioned difficulties and realize the task of tracking a specific person under cross-camera,it is very important for the research of person re-identification.Person re-identification is divided into supervised person re-identification,unsupervised person re-identification,and semi-supervised person re-identification.In recent years,more and more person re-identification tasks rely on variants of GAN for style transfer.Style transfer can not only eliminate the domain variation between cameras but also generate positive samples related to training samples.This paper focuses on style transfer and applies style transfer to one-example person re-identification and unsupervised domain adaptive person re-identification.The work of this paper is as follows:(1)Random style transfer and average label estimation for person re-identification with one-example(RST-AVG).Aiming at the problem of style variations between cameras in one-example task,CycleGAN is used to generate style transfer data.In the training phase,a random style transfer method is proposed to use the original data and style transfer data for training.In the label evaluation phase,an average label estimation method is proposed to use the average characteristics of the original data and the style transfer data to evaluate pseudo labels.The experimental results on the Market-1501 and DukeMTMC-reID datasets show the superiority of RST-AVG.(2)An unsupervised cross-domain person re-identification based on style transfer and camera information(STCF).Aiming at the variation between the source domain dataset and the target domain dataset in unsupervised cross-domain tasks,a two-branch network is proposed.One branch performs supervised training on the source domain data,in which LSR loss is used to reduce model overfitting and provide a basic classification capability for the model;the other branch uses triple loss to perform unsupervised training on the target domain,in which the style transfer data generated by CycleGAN is used as a positive sample and the data filtered by camera information is used as a negative sample.The experimental results on Market-1501,DukeMTMC-reID,and CUHK03 datasets show the superiority of STCF.