Robust Teacher: Self-correcting pseudo-label-guided semi-supervised learning for object detection

Abstract

Semi-supervised object detection methods have become increasingly popular in computer vision owing to their success in reducing data labeling costs. However, low-quality pseudo-labels exhibit difficulties in correction and severely limit the performance of existing state-of-the-art models. To address these problems, based on the contemporary teacher–student dual framework, we develop a novel self-correcting pseudo-label module to generate more reliable predictions for unlabeled data. Simultaneously, to mitigate the inherent class bias in pseudo-labels, we integrate multi-label classification results for measuring the re-balanced focal loss to enable class-agnostic, robust semi-supervised learning. Furthermore, pseudo-label-guided copy-paste is proposed as a simple yet efficient data augmentation technique to enhance instance representation learning within diverse complex scenes. The above key designs constitute our proposed approach, which we call Robust Teacher. Extensive experiments on PASCAL VOC and MS COCO demonstrate that Robust Teacher achieves competitive results and outperforms state-of-the-art models by a large margin. Our comprehensive ablation studies further verify the effectiveness of the key components. The code will be publicly available at https://github.com/Complicateddd/RobustT.