S³MPL:Semi-supervised Semantic Segmentation with Mixed Pseudo Label

Abstract

The introduction of semi-supervised methods reduces the cost of labeling training data. In this regard, semi-supervised semantic segmentation methods have received great attention. In this paper, a consistent regularization method called Semi-supervised Semantic Segmentation with Mixed Pseudo Label (S <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup> MPL) is proposed. In this approach, consistency is applied by feeding the same input image into two segmented networks that are initialized differently and encouraging both networks to make the same predictions. The one with higher confidence in the pseudo one-hot label maps generated by the two segmentation networks is selected as the mixed pseudo label map to supervise the above segmentation networks simultaneously. There are two advantages to this: 1) the quality of pseudo labels is improved; and 2) the segmentation network can not only use divergence between networks to locate its own errors, but also continuously consolidate what it has learned; In addition, our approach also incorporates ClassMix augmentation to mix input images to obtain more diverse data. The experimental results show that our method has achieved competitive performance on Cityscapes and PASCAL VOC 2012.