Unsupervised Semantic Segmentation with Feature Fusion

Abstract

Despite the significant progress in semantic segmentation, its further development is still limited by the requirement for large amounts of annotated data. As unsupervised/self-supervised methods have been successfully applied to image classification tasks, some research has turned to unsupervised semantic segmentation. Unsupervised semantic segmentation methods generally use pre-trained classification networks to extract features as supervision signals, but the richness of information in these features is not well considered. In this paper, we demonstrate the richness issue in features and propose a hierarchical feature fusion and multi-scale fusion method to address it specifically for unsupervised semantic segmentation, which ultimately improves the network's performance on the general semantic segmentation datasets Cityscapes and Coco-Stuff.