domain adaptive semantic segmentation attempts to make satisfactory dense predictions on an unlabeled target domain by utilizing the model trained on a labeled source domain. One solution is self-training, which retrains models with target pseudo labels. Many methods tend to alleviate