Adapting Object Size Variance and Class Imbalance for Semi-supervised Object Detection

Semi-supervised object detection (SSOD) attracts extensive research interest due to its great significance in reducing the data annotation effort. Collecting high-quality and category-balanced pseudo labels for unlabeled images is critical addressing SSOD problem. However, most of existing pseudo-labeling-based methods depend on a large fixed threshold select from predictions teacher model. Considering different classes usually have difficulty levels scale variance distribution imbalance, conventional are arduous explore value sufficiently. To address these issues, we propose an adaptive labeling strategy, which can assign thresholds with respect their “hardness”. This beneficial ensuring high quality easier increasing quantity harder simultaneously. Besides, label refinement modules set up based box jittering guaranteeing localization labels. further improve algorithm’s robustness against make labels, devise joint feature-level prediction-level consistency learning pipeline transferring information model student Extensive experiments COCO VOC datasets indicate that our method achieves state-of-the-art performance. Especially, it brings mean average precision gains 2.08 1.28 MS-COCO dataset 5% 10% labeled images, respectively.