Improving Semi-Supervised Object Detection by ROI-Enhanced Contrastive Learning

Semi-supervised object detection has emerged as a promising paradigm to alleviate the data annotation burden by utilizing a small set of labeled data in conjunction with a larger pool of unlabeled data. Current state-of-the-art methods commonly employ self-training strategies, using pseudo labels to learn from unlabeled data. However, pseudo labels are inherently noisy, particularly in the early stages of training. In this paper, we propose a contrastive learning approach to enhance semi-supervised object detection. Departing from conventional box-level predictions, our method introduces consistency regularization at the feature-level representation. Specifically, we leverage candidate boxes selected by the Region Proposal Network (RPN) for Region of Interest (RoI)-based contrastive learning and introduce pixel-level comparisons for spatial-aware loss calculation. Our experiments demonstrate that the proposed RoI-enhanced contrastive learning effectively enables the model to extract additional information from unlabeled data.