Towards semi-supervised adaptive oriented object detection
Ye Hang Zhu, Yan Bai, Shixin Cen, Cuihong Xue · Remote Sensing Letters · 2025
Semi-Supervised Object Detection (SSOD), which utilizes a small amount of labelled data and a large amount of unlabelled data for training and prediction, has garnered significant attention in recent years. However, existing sampling methods overlook the differentiate of selected positive samples and the distribution characteristics. To address these issues, we propose a novel Semi-supervised Adaptive Oriented Object Detection (SAOD) method, which mainly consists of the Boundary-guided Adaptive Weighting (BGAW) strategy and the Foreground Classification Rebalancing (FCR). The BGAW strategy is composed of the Boundary Fusion Module (BFM) to integrate low-level local edge information and high-level global location information of objects, and the Sample Adaptive Evaluation (SAE), which adopts joint optimization and normalized shape distance to assess the quality of positive samples. Furthermore, a category-related focusing factor for adaptive computation is applied in classification loss, which progressively increases the weight of poorly performing categories during training, effectively mitigating the class imbalance issue. Experimental results on publicly available large-scale remote sensing datasets demonstrate the effectiveness of our proposed method.