Fused-IoU Loss: Efficient Learning for Accurate Bounding Box Regression

The importance of the loss function in object detection algorithms based on deep learning has grown significantly technological progress. The accuracy of object detection is significantly affected by bounding box regression, which is a crucial factor. Since the introduction of the Intersection over Union (IoU) loss in 2016, many improvements have been proposed based on this loss function. These studies considered various geometric factors related to bounding boxes, and constructed penalty terms to address this issue. This paper summarizes these functions and introduces a new Fused IoU (FIoU) loss function that leads to superior performance. The FIoU loss function not only solves the problem of gradient vanishing during the backpropagation process of the IoU loss function but also solves the problem of some IoU-based loss functions degenerating into IoU loss functions under certain conditions. In addition, in the simulation experiments, the FIoU loss function resulted in faster convergence speed. In our ablation experiments across different datasets and algorithms, our aim was to compare the mAP metrics under different loss functions. On the test set of the Pascal VOC dataset, employing the Faster R-CNN algorithm, FIoU demonstrated improvements of 1.1% and 1.7% over GIoU and Smooth
$\ell _{1}$
, respectively. With the YOLOX algorithm, FIoU outperformed GIoU and IoU by 1.0% and 0.8%. Utilizing the YOLOv7 algorithm, we evaluated seven loss functions, achieving optimal results with FIoU. On the validation set of the MS-COCO 2017 dataset, using YOLOv7 and YOLOv8, FIoU exhibited gains of 0.4%, 0.2%, 0.2% over EIoU, DIoU, GIoU, and 0.3%, 0.5%, 0.3% over EIoU, DIoU, GIoU, respectively.