基于困难类别感知数据增强的鲁棒性跨域语义分割研究

In the field of deep learning, it has become a challenging research to mitigate the impact of differences between different scenarios on model performance, and unsupervised domain adaptationon semantic segmentation is one of the classical tasks. Currently, unsupervised domain adaptation methods based on self-training have achieved good results, however, the existence of hard categories makes it difficult to obtain better performance, and the incorrect pseudo-labelin the target domain further hinders the optimization of cross-domain segmentation. To address the above problems, this paper proposes a cross-image hard-awareaugmentation and a Mean-Teacherbased consistency constraint, respectively, to enhance the proportion of hard category in the pseudo-label and improve the robustness of the cross-domain segmentation model. Experiments are conducted on GTA5-to-Cityscapes and SYNTHIA-to-Cityscapes and both objective metrics and visualization results demonstrate the effectiveness of the proposed methods in this paper.