Study of cross-domain person re-identification based on DCGAN

Person re-identification(re-ID) techniques have been rapidly improving with the development of deep neural networks, and the accuracy of fully supervised re-ID models is already very high. However, when person re-identification models with supervised learning are used directly in unlabeled target scenes, the accuracy is greatly reduced. This is due to the large disparity between the domains of different datasets, such as resolution, lighting changes and occlusion. At the same time, the existing re-ID methods based on domain transfer have the problem of blurred images due to big gap between domains. Therefore, in this paper, a DCGAN (Deep Convolutional Generative Adversarial Network) is added to the unsupervised cross-domain re-ID model, which can allow the target domain distribution to be adequately fitted to the labeled domain distribution, so that the re-ID model trained in source domain can be used on the target domain without large fluctuations due to domain shift. By comparing with the UMDL(Unsupervised Multi-task Dictionary Learning) method, rank-1 and rank-5 are improved by 9.61% and 17.81%, respectively, when the model trained from Market-1501 is used in the GRID dataset, and by 9.52% and 14.14%, respectively, when compared with the GAN(Generative Adversarial Networks)-based method under the same experiment.