Improved Design Based on IoU Loss Functions for Bounding Box Regression

Abstract

In object detection, the bounding box regression loss calculation has a great influence on the positioning effect of object detection. At present, the common loss function is smooth L1 loss or Intersection over Union(IoU)-based loss function. For the problems of large training error and low training accuracy based on the IoU loss function, two improved versions of loss functions BaIoU and BhIoU are proposed. Among them, BaIoU combines the balanced L1 loss and the original IoU loss function; based on BaIoU, BhIoU increases the loss gradient of IoU by improving the form of IoU to improve the algorithm of IoU. The bounding box regression simulation experiment proves that BaIoU and BhIoU can effectively overcome the problems of slow convergence and large training error of the loss function based on IoU. Using the MS COCO dataset, training tests are conducted using a two-stage object detector and a single-stage object detector, and the test results prove that BaIoU and BhIoU can improve the performance of the object detector with better localization accuracy than the existing IoU-based loss functions.