Abstract
Nowadays, amount of computer vision tasks achieve better performance by using Generative Adversarial Nets(GAN), the main and basic purpose of GAN is to generate unlabeled images so that training deep learning models with these images can hopefully boost system performance significantly. In this paper, we propose a specific method of properly making use of unlabeled images generated by GAN for person re-identification(re-ID) baseline training, we propose label boosting regularization for outliers(LBRO) algorithm, which assigns reasonable label distributions to different numbers of unlabeled images when computing loss. We verify that LBRO can reduce overfitting as well as underfitting under various circumstances, also greatly enhances the performance in person re- ID baseline. Experiments on three mainstream datasets: Market-1501, CUHK03, DukeMTMC-reID show effective results of our method with well-tuned hyperparameters. We can conclude that adding more unlabeled images into training set as a form of regularization can combat overfitting but one must consider careful label distributions of different sizes of GAN images when computing loss.