Region-Guided Pixel-Level Label Generation for Weakly Supervised Semantic Segmentation
Xinyu Fu, Xiao Lan Yao · 2021
The lack of reliable segmentation labels is the major obstacles to weakly supervised semantic segmentation. We provide a pseudo-label generation approach based on a deep convolutional neural network, which is supervised by the image-level category labels only. However, the limitation of the region of interest in the targets influences the effectiveness and integrity of the traditional methods in obtaining pixel-level mask annotations. This paper studies the characteristics of class activation mapping in classification network, focusing on the methods to enhance the localization ability of class activation mapping. We propose a Region-guided Pixel-label Generation framework (RPG) for semantic segmentation. The proposed region guidance mechanism decreases the influence of category supervision and makes use of the known high-level semantic information to guide the network, attaining more complete pixel-level annotations via expanding the regions of interest. Experimental results of training and validation on the PASCALVOC 2012 data set prove to achieve better pixel labeling and segmentation accuracy comparing with state-of-the-art methods.