Semi-Supervised Object Detection with Adaptive Pseudo-Label Mining

To address the problem of threshold selection and scale imbalance in semi-supervised object detection, we presents an adaptive pseudo-labeling mining method to tackle the aforementioned issues. Firstly, the pseudo-labels were divided into three types: negative pseudo-labels, uncertain pseudo-labels and positive pseudo-labels, and the Gaussian mixture model was used to set a threshold for each type of pseudo-labels, and the pseudo-label threshold was dynamically adjusted to solve the problem of fixed threshold. Then, the negative samples in the uncertain pseudo-labels are further identified and mined by the adaptive pseudo-label mining module to extract more high-quality pseudo-labels for model training. In order to resolve the issue of scale imbalance, this paper proposes a strategy combining dual Region Generative Network (DRPN) and Upsampling based on feature Content-aware reorganization (CARAFE). The experimental findings demonstrate that the mAP value of the proposed method on the 5% and 10% labels of the Microsoft COCO dataset is increased by 0.5 and 1.1, respectively. This outcome validates the efficacy and superiority of the proposed method in the domain of semi-supervised object detection.