The Research On Unsupervised Domain Adaptation Algorithm For Person Re-identification

In the field of computer vision,person re-identification(re-ID)is increasingly important because of its high efficiency in criminal investigation and intelligent video surveillance.Therefore,re-ID is a significant task of high research value.Re-ID aims at retrieving images of a person from a large-scale gallery collected by multiple cameras,given a query of person-of-interest.In recent years,although the re-ID performance has been significantly boosted,but re-ID model trained on one dataset usually cannot work effective on another one.In order to solve the problem of domain gaps between multiple datasets in re-ID,this paper proposes a domain adaptation person re-identification algorithm based on CycleGAN.Besides,in order to solve the problem of person identity label loss during unsupervised image translation,a DTGAN-based domain adaptation person re-identification algorithm is proposed in this paper.It is proved in experiments that this new structure work well to solve the problem of adaptability of re-ID model on different datasets.The main work of this paper is as follows:(1)this paper proposes a baseline method based on CycleGAN image-to-image translation to solve the problem of domain adaptation in re-ID.In this paper,the method is divided into two steps.first,the source images are translated in an unsupervised manner,and the translated images retain the style of target images and the ID labels in source domain;second,the translated images are used as the training data for supervised feature learning and the model is finally trained on target dataset.Experiments show that this baseline method solve the problem of domain gaps in re-ID and has a significant improvement in re-ID accuracy compared with the direct transfer method.(2)This paper proposes a novel unsupervised learning network,named DTGAN(Double-domain Translation Generative Adversarial Network),to improve the baseline method.DTGAN is composed of two parts,one is the Siamese network and the other part is CycleGAN.the CycleGAN learns the generative mappings between source domain and target domain in the process of image-to-image translation.Siamese network pulls close translated image and its counterpart in source domain.On the contrary,because the translated images should maintain different ID information from any target images,the Siamese network pushes them further away.Therefore,the ID information can be retained in the process of image-to-image translation.Finally,we employ the strategy of label smoothing regularization(LSR)to eliminate the noise generated in the process of model training.Experiments show that DTGAN can generate pedestrian images of high quality and solve the problem of domain gaps in re-ID.In addition,DTGAN achieves competitive re-ID accuracy on Market1501 and DukeMTMC-reID datasets.