GuidedMix-Net: Semi-supervised Semantic Segmentation by Using Labeled Images as Reference

• DOI: 10.1609/aaai.v36i2.20137
• Publication Date: 2022-07-04T10:26:29Z
• Authors: Peng Tu; Yawen Huang; Feng Zheng; Zhenyu He; Liujuan Cao; Ling Shao
• Keywords: Pascal (unit); Labeled data; Transfer of learning; Supervised Learning

Semi-supervised learning is a challenging problem which aims to construct model by from limited labeled examples. Numerous methods for this task focus on utilizing the predictions of unlabeled instances consistency alone regularize networks. However, treating and data separately often leads discarding mass prior knowledge learned In paper, we propose novel method semi-supervised semantic segmentation named GuidedMix-Net, leveraging information guide instances. Specifically, GuidedMix-Net employs three operations: 1) interpolation similar labeled-unlabeled image pairs; 2) transfer mutual information; 3) generalization pseudo masks. It enables models can higher-quality masks samples data. Along with supervised data, prediction jointly generated mixed Extensive experiments PASCAL VOC 2012, Cityscapes demonstrate effectiveness our achieves competitive accuracy significantly improves mIoU over 7$\%$ compared previous approaches.