The state of the art in semantic segmentation is steadily increasing in\nperformance, resulting in more precise and reliable segmentations in many\ndifferent applications. However, progress is limited by the cost of generating\nlabels for training, which sometimes requires hours of manual labor for a\nsingle image. Because of this, semi-supervised methods have been applied to\nthis task, with varying degrees of success. A key challenge is that common\naugmentations used in semi-supervised classification are less effective for\nsemantic segmentation. We propose a novel data augmentation mechanism called\nClassMix, which generates augmentations by mixing unlabelled samples, by\nleveraging on the network's predictions for respecting object boundaries. We\nevaluate this augmentation technique on two common semi-supervised semantic\nsegmentation benchmarks, showing that it attains state-of-the-art results.\nLastly, we also provide extensive ablation studies comparing different design\ndecisions and training regimes.\n