Diagonal-Angle-Foreground IoU Loss Function for Small Object Detection

Boundary box regression is a crucial step in object detection, directly influencing the model's ability to accurately locate objects. Traditional Intersection over Union (IoU) loss functions struggle to provide effective gradient information when dealing with the situation where the predicted box and the ground truth box's center point coincide, leading to slower convergence. In the small objects detection, due to their smaller proportion in the image, it is essential to fully utilize their information to improve boundary box regression. Addressing these issues, this paper proposes a loss function based on diagonal, angle, and foreground information, named Diagonal-Angle-Foreground IoU (DAFIoU) loss. The use of diagonal and angle information effectively resolves the problem of the loss function failing to provide gradient information. Simultaneously, incorporating foreground information as a loss term significantly enhances the information utilization of small objects. This paper validates the effectiveness of the proposed loss function through simulated experiments. Additionally, using YOLOv8 and Faster R-CNN as the bench-mark detection algorithm, comparative experiments between IoU series losses and DAFIoU losses are conducted on the VisDrone2019 and SODA-D10 small object public datasets. The experimental results demonstrate the excellent performance of the proposed loss function across different datasets and algorithms.