Object detection in aerial images, aiming to identify objects of interest on the earth’s surface, has attracted increasing attention for its broad application prospects. However, annotating massive training data in aerial images is challenging due to the high labour costs. Semi-supervised object detection (SSOD) is a promising method to tackle this issue by training detectors with limited labeled and massive unlabeled data. This paper presents a cost-effective SSOD method for aerial images called ODAdapter, which comprises an adaptive pseudo-label filtering module and a novel consistency loss function. Specifically, the filtering module dynamically filters pseudo-labels by calculating the confidence density as an adaptive threshold, enriching the diversity of pseudo-labels. The consistency loss function regularizes the orientation and aspect ratio differences between each pair of pseudo-labels and predictions, improving the accuracy of pseudo-labels. We examine the performance of our method under various experimental settings on three challenging datasets: DOTA-v1.5, DIOR-R, and UCAS-AOD. Our ODAdapter achieves 67.86 mAP on the DOTAv1.5 benchmark, surpassing the supervised baseline by +2.44 mAP.