UCR-SSL: Uncertainty-Based Consistency Regularization for Semi-Supervised Learning

Abstract

Recently, semi-supervised learning methods are being actively developed to increase the performance of neural networks by using large amounts of unlabeled data. Among these techniques, pseudo-labeling methods have the advantage of low computational complexity, but are vulnerable to missing annotations. To solve this problem, we propose a method called uncertainty-based consistency regularization (UCR). UCR models a detection head to obtain different outputs for input images and computes a feature map of each. Subsequently, these feature maps are matched with the original and filtered ground truth (GT), and are classified as positive and negative samples, respectively. In this process, missing samples are generated by the filtered GT; therefore, we use a specialized loss function designed to reduce the logit difference of the samples for robustness against missing annotations. We also use the uncertainty extracted through Gaussian modeling as a criterion for annotation filtering to train the network to focus on reliable results. As a result of experiments with an SSD model on the Pascal VOC dataset, the proposed approach achieved an improvement of 0.7% in terms of mAP compared to a baseline method.