Appendixof" Semi-supervisedSemanticSegmentation withPrototype-basedConsistency Regularization "

Next, we provide another two ablation studies to further inspect our approach. In the main paper, the visualization of feature distribution in Figure 2 (c) has demonstrated that our approach can encourage a more compact within-class feature distribution and thus ease the largeintra-class variation problem inthesemi-supervised semantic segmentation. Yellowdotted boxeshighlight thesegments where our method performs better than the comparison method, i.e., our method can better perceive the boundaryofobjects. Cutmix: Regularization strategy to train strong classifiers with localizable features.