Semi-Supervised Semantic Segmentation Constrained by Consistency Regularization

In this paper, we propose a self-training based method for semi-supervised semantic segmentation. Our method utilizes k perturbed images of each unlabeled image to generate a new mask through the proposed vote operation which in turn is used as a supervision signal to train the model. The k predicted masks of perturbed images can provide nontrivial knowledge that is not captured by a single prediction and the proposed vote operation enables the model to output a low entropy prediction. Consistency regularization is applied between generated mask and k masks in terms of MSE loss in our network. Extensive experiments are conducted on different datasets to evaluate the effectiveness of our method. We show that the proposed method surpasses previous state-of the-art semi-supervised methods on ISIC 2017 dataset and achieves competitive performance on PASCAL VOC 2012 dataset.