MLCB-Net: a multi-level class balancing network for domain adaptive semantic segmentation

To solve the long-tail distribution problem in domain adaptive semantic segmentation, we propose a novel multilevel class balancing network (MLCB-Net). We adapt a novel frequency fusion module (FFM) by using prior knowledge to guide the domain adaptive semantic segmentation network, thus carrying out regular constraints in global training. Furthermore, in the domain adaptation process, we introduce a dual-branch balancing module (DBBM) to resample the class-level features, which makes the model not only improve the sensitivity to low-frequency classes but also does not damage the feature representation ability of the classifier. In addition, we combine self-supervised learning strategies with our proposed modules to further improve segmentation performance. Experiments on two baseline tasks, GTA5 to Cityscapes and SYNTHIA to Cityscapes, show that MLCB-Net achieves a new state-of-the-art benchmark and is more robust.