Semi-supervised Semantic Segmentation with Complementary Reconfirmation Mechanism

Pseudo-labeling based methods are commonly employed for effectively utilizing unlabeled data in semi-supervised semantic image segmentation. It tends to select high-confidence pixels in images as pseudo labels and discard most of the pixels predicted with low confidence, but each pixel is valuable for accurate segmentation. Therefore, we propose a semi-supervised semantic image segmentation algorithm based on complementary reconfirmation mechanism (CR-Seg) to constrain the low-confidence pixels. Firstly, the predictions are divided into high-confidence and low-confidence pixels by a dynamic threshold. The High-confidence pixels supervise the predictions of the student model and guide the classifier to learn what they are. While the low-confidence pixels are equally important for model training, they are used to generate complementary reconfirmation loss for providing additional information as complementary labels. Our method achieves mIoU of 69.12, 73.84, 74.03, and 76.91% under 1/16, 1/8, 1/4, and 1/2 partitions of the classic PASCAL VOC 2012. The experimental results demonstrate that low-confidence pixels can provide more information to the model as complementary labels, thereby improving the model’s segmentation performance.