Semi-Supervised Active Learning for Object Detection

Abstract

Behind the rapid development of deep learning methods, massive data annotations are indispensable yet quite expensive. Many active learning (AL) and semi-supervised learning (SSL) methods have been proposed to address this problem in image classification tasks. However, these methods face a new challenge in object detection tasks, since object detection requires classification as well as localization information in the labeling process. Therefore, in this paper, an object detection framework combining active learning and semi-supervised learning is presented. Tailored for object detection tasks, the uncertainty of an unlabeled image is measured from two perspectives, namely classification stability and localization stability. The unlabeled images with low uncertainty are manually annotated as the AL part, and those with high uncertainty are pseudo-labeled with the detector’s prediction results as the SSL part. Furthermore, to better filter out the noisy pseudo-boxes brought by SSL, a novel pseudo-label mining strategy is proposed that includes a stability aggregation score (SAS) and dynamic adaptive threshold (DAT). The SAS aggregates the classification and localization stability scores to measure the quality of predicted boxes, while the DAT adaptively adjusts the thresholds for each category to alleviate the class imbalance problem. Extensive experimental results demonstrate that our proposed method significantly outperforms state-of-the-art AL and SSL methods.