Semi-Supervised Active Learning for Object Detection

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