Generative Adversarial Network Based Person Re-identification Method

Person re-identification(re-ID)is the basis of person tracking and Intelligent Safe-guard Systems.It is a key method of building Safe City and Intelligent City.Person re-ID is facing two problems:the lack of pose diversity in dataset and the domain gap between RGB images and infrared images.To overcome them,we implement two ge-nerative neural networks for pose transfermation and domain transfermation.Based on the distribution of generated data,we propose a Triplet Adversarial Loss to train the generative model and re-ID model together.Under adversarial training,we could get a better re-ID model.Experimental results on public datasets show that our method can improve the re-identification accuracy,the robustness against pose variations as well as the cross-domain retrieval ability.The main achievement of this paper is:1.We implement a pose transfer network and use the generated images to boost person re-ID training dataset.This network can change the pose in pedestrain images with realistic results.By putting pose transferred images into training da-taset,pose diversity can be boosted.Experimental results on Market-1501 dataset show that this data augmentation method can achieve a re-ID model with leading robustness and high accuracy.2.We propose an adveresarial training strategy for generative models and re-ID models.Identities should be kept for generated pedestrain images.Based on this principle,we propose a Triplet Adversarial Loss.Experimental results on public datasets show that with this loss function,re-ID model could improve its identification accuracy and robustness.3.We propose a method for RGB-infrared cross domain images retrieval.We im-plement a domain transfer network which can convert a RGB image into infrared image.It is trained with the proposed Triplet Adversarial Loss.When testing,we propose to extract a cross-domain feature and a infrared-domain feature se-perately for each image.The two features are combined to get the final feature.Experimental results show that our method can achieve great RGB-infrared cross domain retrieval performance improvement and outperforms the state-of-the-art methods.