Class-Aware Feature Regularization for Semantic Segmentation.

In this paper, to address the problem of intra-class consistency and inter-class variation in the deep convolutional neural network (CNN) based methods for semantic segmentation of images, we propose a class-aware feature regularization strategy to revise the features extracted by a deep convolutional neural network, without any change of the original network structure. A pixel-context similarity term is proposed to measure the consistency between feature vectors of pixels and class centers, which guarantees the intra-class consistency of pixels in the interior of an object and is supervised by a One-Hot label to preserve the inter-class variation of different objects. Based on the similarity term, we design a lightweight and efficient plug-in loss term to ensure that the features yielded by a deep CNN possess the quality of intra-class consistency and inter-class variation. As our ideal can be fulfilled effectively by the proposed plug-in loss term, we can simply incorporate it into a CNN-based segmentation model without changing the model structure. The effectiveness of the proposed strategy is proved by incorporating the loss term into some state-of-the-art segmentation models on Cityscapes and ADE20K datasets.