DRANet: Dual-threshold Guided Reliability Aware Network for Semi-Supervised Image Semantic Segmentation

Abstract

Different from supervised semantic segmentation task, semi-supervised semantic segmentation (SSSS) aims to alleviate the burden of time-consuming pixel-wise manual labeling. Although existing methods have achieved the promising performance with a small amount of labeled images, they still suffer from an insufficient mining on a large amount of unlabeled images due to simple evaluation way for them. To address this issue, we propose a novel Dual-threshold Guided Reliability Aware Network (DRANet) for SSSS task. To comprehensively evaluate the reliability of unlabeled images, we provide a dual-threshold guided augmentation module (DGAM), which not only introduces two predefined thresholds to group unlabeled images with different confidences according to their learning status, but also performs different data augmentations on unlabeled images based on their evaluation results. In particular, for an unlabeled image with higher confidence, a double-stream augmentation strategy is adopted, leading to a more sufficient learning for it. Experiments on Cityscapes and Pascal VOC 2012 datasets confirm the superiority of our DRANet over existing methods. Code is available at https://github.com/ZY-IMU-CV/DRANet_QLZ_2024.