Learning an unbiased classifier from imbalanced image datasets is achallenging task, since the classifier may strongly bias towards majorityclasses. To address this issue, some deep generative models-basedoversampling methods have been proposed. However, most methods paylittle attention to the decision boundary, which may contribute tiny tolearning an unbiased classifier. In this paper, we focus on the decisionboundary and propose a similar classes latent distribution modellingbasedoversampling method.Specifically, first, we model each class asdifferent von Mises-Fisher distributions, thereby aligning feature learningwith the class distributions. Furthermore, we develop a distanceminimization loss function, which makes similar classes closer in latentspace. The generator can learn more shared latent features from thedecision region. In addition, we propose a boundary sampling strategy,which uses latent variables between similar classes to generateboundary samples for data balancing. Experiments on four imbalancedimage datasets show that the proposed method achieves promisingperformance in terms of Recall, Precision, F1-score and G-mean.