Abstract: Massive high-quality annotated data is required by fully-supervised learning, which is difficult to obtain for image segmentation since the pixel-level annotation is expensive, especially ...