ALS Teacher: Active Label Selection for Semisupervised Oriented Object Detection in Remote Sensing Imagery

Semisupervised object detection (SSOD) in remote sensing scenarios is fundamentally challenged by orientation-sensitive pseudolabel noise under limited supervision, where arbitrary object rotations and complex spatial distributions significantly degrade label reliability. Although SSOD has shown promising results in natural images, existing methods predominantly focus on horizontal object detection and struggle to generalize to oriented settings. In this article, we present ALS Teacher, a semisupervised framework specifically designed for oriented object detection within a two-stage detector. To improve the reliability of pseudolabels under limited supervision, we introduce an active label selection strategy (ALSS), which identifies high-value labeled samples using comprehensive selection metrics. Furthermore, to enhance orientation-aware representation learning from unlabeled data, we develop orientation-preserving interpolation augmentation, a geometry-consistent interpolation augmentation strategy designed to preserve spatial coherence and rotational characteristics. In addition, we propose the global oriented loss function, which incorporates global spatial distributions into oriented regression, fostering more stable and expressive orientation modeling in the student detector. Extensive experiments on DOTA-v1.5 and CODrone benchmarks demonstrate the effectiveness of our framework. With only 10% labeled data, our method achieves 52.26% mAP on DOTA-v1.5 and 37.84% mAP on CODrone, surpassing the supervised Oriented R-CNN baseline by 7.78% and 4.37% mAP, respectively. These results highlight the potential of our approach as a scalable solution for semisupervised oriented object detection in complex remote sensing imagery.