A Residual Correction Approach for Semi-supervised Semantic Segmentation

Fully-supervised deep learning models have achieved a great success in complex semantic segmentation tasks. However, the segmentation annotations are prohibitively expensive, which causes a growing interest in the methods that require lower annotating cost but still achieve a competitive performance. This paper proposes a residual correction approach based on self-training for semi-supervised semantic segmentation. We train a residual correction network built on top of the segmentation network with labeled data to predict a residual of the original segmentation. For unlabeled data, the output of the residual correction network is combined with the original segmentation to form the pseudo label used to train the segmentation network. Extensive experimental results on the PASCAL VOC 2012 and the Cityscapes datasets demonstrate the effectiveness of the proposed approach.