Mining Oriented Information for Semi-Supervised Object Detection in Remote Sensing Images

Traditional object detection requires extensive annotation, consuming considerable time and manpower. In recent years, semi-supervised object detection (SSOD) methods, which utilize a blend of unlabeled and labeled data to train object detectors, have been extensively researched. SSOD has made significant progress and can achieve similar levels of accuracy as the fully supervised methods. However, existing SSOD approaches primarily focus on horizontal objects in natural scenes, with scant research on other scenarios such as remote sensing images. This paper proposes a novel semi-supervised oriented object detection algorithm based on a two-stage object detector. For oriented objects in remote sensing, we particularly emphasize the utilization of oriented information of remote sensing objects to generate precise pseudo-label and improve the learning capability of the student network. Iterative labeled data filtering is performed by incorporating metrics with designed measurements. Valuable annotated samples can enhance the quality of pseudo-label generation. A strong augmentation method has been designed to utilize rotational information, enabling the student network to learn more diverse features. Additionally, we investigate the long-tail distribution problem in remote sensing images and mitigate the bias brought by category imbalance through phased training and post-processing in detection. Our experiments demonstrate that the designed semi-supervised oriented object detection method surpasses existing methods in the DOTAv1.5 benchmark, culminating in state-of-the-art performance.

I. Introduction

In the realm of fully supervised object detection, the availability of adequately labeled data is essential. The process of labeled data is both time-consuming and costly. Recently, the SSOD approach has garnered significant interest in the field. This emerging technique enhances the training of object detectors by integrating labeled data with more readily accessible unlabeled data. Most SSOD methodologies focus on detecting objects using horizontal bounding boxes in traditional settings. However, in more complex scenarios, such as the analysis of remote sensing images, objects often require annotation with oriented bounding boxes. Given the high costs associated with labeling oriented boxes, the exploration and development of semi-supervised methods tailored for such contexts are increasingly important. These efforts could substantially reduce costs while improving the accuracy and efficiency of oriented object detection.