A Survey of Weakly-supervised Semantic Segmentation

Abstract

Weakly-supervised image semantic segmentation is a popular technology in computer vision and deep learning today. The main goal of weakly-supervised semantic segmentation is to train a model by images with only coarse or sparse annotations. Specifically, it assigns a label to each pixel through coarse label refinement or sparse label propagation, etc. The existing semantic segmentation has a wide range of applications, which includes pedestrian detection, autonomous driving, medical image segmentation, etc. However, fully-supervised semantic segmentation requires pixel-level annotation, which is expensive in manpower and time, and more and more works have focused on weakly-supervised semantic segmentation in recent years. Thus, this paper provides a review of weakly supervised semantic segmentation. Firstly, this paper summarizes the state-of-the-art research results of weakly-supervised semantic segmentation. Secondly, the widely-used datasets and semantic segmentation models are introduced. Finally, this paper analyzes the existing problems and future development directions in the field of weakly-supervised semantic segmentation.