Person Re-identification Data Augmentation Based On GAN

Person Re-Identification(re-id)aims to search for a specific person in a large-scale short-term scene.However,compared with other computer vision tasks,it is more difficult to ob-tain person data due to privacy and annotation costs.This makes the effect of the person re-identification model largely restricted by the size of the dataset.Data augmentation based on Generative Adversarial Networks(GAN)can provide additional data for person re-identification tasks to improve the performance of the network.This thesis discusses the augmentation method of person re-identification data based on GAN.The second chapter of this thesis systematically analyzes the method of using camera style transfer to augment person data,and conducts experiments on fully supervised person re-identification and unsupervised domains adaptation for person re-identification.Based on a high baseline level,the conclusions drawn are contrary to most of the same types of methods:for fully supervised tasks,there is no need to use camera style transfer or other data augmenta-tion methods to improve performance? for person re-identification tasks with incomplete labels,such as unsupervised domain adaptation for person re-identification,camera style transfer can bring huge performance improvements,but there is no need to rely on specific loss functions,such as identity mapping loss,MS-SSIM,and perceptual loss to constrain person identity? even the loop verification module of Star GAN can be removed to further reduce the training cost.The third chapter of this thesis proposes ME-Star GAN,a data augmentation method that for unsupervised domain adaption person re-identification for dynamic changes of target datasets.In this scenario,Star GAN cannot cope with low training costs.We designed ME-Star GAN with almost no impact on network performance.ME-Star GAN generators can be divided into encoders and decoders.By enhancing the encoder's ability to extract person information and attribute the parameters that affect the generated picture style to the decoder,ME-Star GAN has successfully implemented the multiplexing of the encoder.The parallel training method is used to further improve performance,greatly reduce training costs,and improve the efficiency of dynamically added camera style conversion and actual applications.Different from the existing data augmentation methods that focus on camera style transfer,Chapter 4 of this thesis proposes View Edit GAN for view editing of person images.View Edit-GAN relies on the interpretability principle of the GAN.First,the eigenvectors of person are initially located through the encoder and decoder,and then the vector error is corrected by the cyclic correction algorithm,and finally the hyperplane obtained by the SVM algorithm is used to interpolate the vector.To achieve the editing of the person's perspective,View Edit GAN can edit the perspective of the person in the picture without relying on the pedestrian category,which broadens the way of data augmentation.This thesis summarizes the existing person data augmentation algorithms based on camera style migration,analyzes the effectiveness of different loss functions under high performance?and proposes ME-Star GAN to solve the problem for practical application scenarios? finally jumps out of the camera Style transfer,and proposed View Edit GAN that can augment pedestrian data through editing perspectives.