Survey of weakly supervised semantic segmentation methods

Weakly supervised image semantic segmentation is quite popular in computer vision and machine learning today. The existing supervised semantic segmentation approaches require lots of pixel-based annotation for training, which are labor- and time-consuming to obtain. By comparison, weakly supervised methods don't need the pixel-level label, and they can do the semantic segmentation work by only using the image-level label. This paper provides a review on weakly supervised semantic segmentation methods. First, the existing methods are described. Then, the main datasets and evaluating indicator are exposed. Subsequently, the results are discussed. Finally, the conclusions are given.