Optimization for Anchor-Free Object Detection via Scale-Independent GIoU Loss

Arbitrary-oriented target detection is widely used in optical remote-sensing image processing, and there have been lots of anchor-based detectors using horizontal bounding boxes. However, the image targets of various scales and shapes make it difficult to tune optimal anchor parameters, whereas the complex background and nonmaximum suppression (NMS) require well-aligned bounding box to predict dense targets. In this letter, a scale-independent IoU (SIoU) loss is proposed for bounding box regression, which can adaptively adjust the shape of predicted boxes and speed up the convergence. Besides, the regression branch of the fully convolutional one-stage object detector (FCOS) is refined to implement the novel intersection over union (IoU) loss for rotated bounding box regression. Extensive experiments on HRSC2016 and a large-scale dataset for object detection in aerial images (DOTA) show that our method obtains 88.1% mean average precision (mAP) under an IoU threshold of 0.5 on HRSC2016, which is 1.1% higher than generalized IoU (GIoU) loss and 0.7% than complete IoU (CIoU) loss.