Absolute size IoU loss for the bounding box regression of the object detection

Abstract

With the wide application of deep learning in the field of computer vision, the technology of object detection continues to make breakthroughs, and the bounding box regression technology is closely related to the accuracy of object detection results. This study proposes an Absolute size IoU (AIoU) loss function for bounding box regression, which further improves the object detection accuracy. Firstly, this study introduces the common location loss functions in bounding box regression, and then describes the limitations of common loss functions based on Intersection over Union (IoU). To overcome these limitations, this study puts forward an AIoU loss function, which can facilitate bounding box regression. Specifically, when the loss penalty term becomes invalid, it can replace the existing penalty term for model optimization. In addition, it can focus the models further on the difficult objects during training. Moreover, as a comprehensive regression factor, this penalty term contains various optimization features. The effectiveness and wide range of application of the AIoU proposed are demonstrated in experiments with three different detectors. It improves the performance of YOLOv4 by 0.61% mAP on the VOC dataset and by 1.98% mAP on the COCO dataset. Finally, we have obtained α-AIoU which uses a power function for AIoU improvement, and it achieves the best performance in the experiments. The evaluation results on several different detectors show that the method proposed in this study has important application significance for object detection technology. The code related to the AIoU loss experiment is at https://github.com/tiandii/AIoU-loss-experimental-code-.git.