A Shape-Based Quadrangle Detector for Aerial Images

The performance of oriented object detectors has been adversely impacted by the substantial variations in object orientation. In this paper, we propose a simple but efficient object detection framework for oriented objects in aerial images, termed QuadDet. Instead of adopting oriented bounding box to represent the object, we directly predict the four vertices of the object’s quadrilateral. Specially, we introduce a fast sorting method for four vertexes of quadrangles, called the Vertex Sorting Function. The function confirms that the vertexes can compose a valid quadrangle by sorting tangents of the vertexes. Furthermore, we employ an efficient polygon IoU loss function, named the PolyIoU Loss Function, to progressively align the predicted quadrangle’s shape with the ground truth. Under these strategies, our model achieves competitive performance. Without bells and whistles, our method with ResNet50 achieves 73.63% mAP on the DOTA-v1.0 dataset running at 23.4 FPS, which surpasses all recent one-stage oriented object detectors by a significant margin. Moreover, on the largest dataset DOTA-v2.0, our QuadDet with ResNet50 obtains 51.54% mAP. The code and models are available at https://github.com/DDGRCF/QuadDet .