GRDet: Rotated Object Detection in Remote Sensing Images Based on Gaussian Distribution

In recent years, many rotated object detection (ROD) methods have been proposed and have attracted wide attention in many fields. Most of them use anchor-based or Gaussian heatmaps for label assignment (LA), which cannot capture the shape and orientation characteristics of the rotated object and introduce a large number of hyper parameters. At the same time, most methods only add angle regression or use enclosing rectangles to realize ROD, which cannot express the object well. In this paper, we propose a new method for ROD, named GRDet, which is keypoint-based Anchor-free algorithm. GRDet can adaptively learn and represent an object with point sets, discarding the limitation of anchor on the size and orientation of the object. Specifically, we introduce a conversion function that is able to transform the point set into a rotated bounding box (RBB) for precise localization and classification. In addition, we propose a Gaussian-based dynamic label assignment (GDLA) strategy to realize the assignment of positive and negative (P&N) samples, which can adaptively learn according to the size and orientation characteristics of any rotated object. Moreover, we define an intersection over union (IoU) suitable for ROD, called Gaussian-IoU, which simulates the calculation of IoU by Gaussian distribution and solves the case that some points are not differentiable. Furthermore, we design a dynamic spatial quality constraint (DSQC) for RBB, which can dynamically evaluate the quality of the predicted RBB, and adaptively select high quality RBB. We use KFIoU loss and introduce Gaussian center loss to supervise the training of the network. Extensive experiments with DOTA dataset demonstrate the effectiveness of our proposed method.