Weakly Supervised Rotated Object Detection Based on Multi-Scale Attention Mechanism

In the field of remote sensing object detection, the annotation work of rotated bounding boxes is extremely tedious and consumes a large amount of human and time costs. To effectively alleviate this problem, this paper proposes a novel weakly supervised rotated object detection method specifically designed for remote sensing scenarios. This method only requires horizontal bounding box annotations for training, which significantly reduces the annotation requirements. The core of this method is a feature fusion network that integrates a multi-scale receptive field module and a Global Attention Mechanism (GAM), aiming to enhance the model’s feature representation ability for multi-scale and multi-directional objects. We design a dual-branch learning framework: a weakly supervised branch associates horizontal box annotations with rotated box predictions through geometric constraints to learn and generate preliminary rotated detection results; a self-supervised branch applies rotation transformations to the original image and uses consistency loss to constrain the prediction outputs of the two branches, thereby providing a reliable supervision signal for the estimation of the object’s orientation angle. This design ingeniously solves the core challenge of angle learning ambiguity in the weakly supervised setting. Extensive experiments on two mainstream remote sensing detection benchmarks, DOTA and DIOR-R, show that the proposed method performs excellently. It not only significantly outperforms other weakly supervised comparison methods but also is comparable to or more competitive than many fully supervised methods that require full rotated annotations. At the same time, this method also shows great advantages in memory efficiency and inference speed. The results show that the method proposed in this paper provides an effective way to achieve efficient and low-cost remote sensing rotated object detection.