DTFSeg: A Dynamic Threshold Filtering Method for Semi-Supervised Semantic Segmentation

Recently, the research on semi-supervised semantic segmentation has made rapid progress, where a large number of unlabeled images with pseudo labels are adopted for boosting performance. Despite their achievement, how to get high-quality pseudo labels still remain challenging. Most methods would use complexly designed threshold strategies for pseudo tag generation. In this article, we propose a semi-supervised semantic segmentation method based on simple threshold filtering and self-training. In the process of generating pseudo-labels, the method deals with the thresholds of different categories of image pixels separately. It filters the labels of each category of pixels by dynamically changing thresholds to guide the model to train. This method is a general strategy and can be combined with the existing semi-supervised semantic segmentation methods based on generating pseudo-labels. We fully demonstrate its effectiveness on the Cityscapes dataset and UAVid dataset.