<p>Unsupervised Domain Adaptation (UDA) intends to transfer the knowledge learned from labeled source domain to unlabeled target domain. Most existing methods employ domain adversarial training to align the feature space distributions of two domains. However, these methods may destroy the discriminative structural information. In this paper, we propose a Dynamic Prototype-guided Structural Information Maintaining (DPSIM) approach to preserve the structural information of the target domain based on pairwise semantic similarity. Specifically, we propose a dynamic prototype learning module to learn the categorical intrinsic representation of the source domain and then to predict the similarity of pairwise samples of the target domain. Finally, a structural information maintaining module is proposed to restrict the target domain by discriminating structural information. Extensive experiments on both image classification and object detection tasks demonstrate the effectiveness of our method.</p>