Aspect-Ratio Aware Dynamic Label Assignment for Oriented Object Detection

Due to the presence of objects with large aspect ratios and variations in their angles, the use of oriented object detectors for remote sensing images faces significant challenges. As a result, detection strategies for oriented object detectors have been extensively studied. The label assignment strategy and angle loss calculation often need to consider the objects’ shape characteristics fully within these detectors. A small deviation in the angle prediction leads to a dramatic drop in intersection over union (IoU) between the prediction box and the ground-truth (GT) box. This can lead to problems with large aspect ratio objects in the presence of insufficient sampling and loss misalignment. To address these issues, this letter proposes a new and flexible aspect-ratio aware (ARA) dynamic label assignment strategy and aligned angle loss (AAL) strategy for oriented object detection. Precisely, we utilize information such as the prediction classification score, the IoU between the prediction boxes and the GT boxes, and the aspect ratio of the GT boxes to evaluate the quality of the candidate samples to dynamically select positive and negative samples to improve the quality of the positive samples. In addition, to further enhance the detection capability for objects with large aspect ratios, we decouple the localization loss and incorporate aspect ratio information into the loss calculation, thereby aligning angle loss for objects with varying aspect ratios. Extensive experiments on popular remote sensing datasets, such as DOTA and HRSC2016, validate the effectiveness and superiority of our proposed method.