Rethink the IoU-based loss functions for bounding box regression

Abstract

The $\ell_{\mathrm{n}}-$norm loss is widely used as the bounding box regression loss function in the existing object detectors, but Intersection over Union (IoU) is used to evaluate the accuracy of the bounding box locating. There is no inevitable correlation between low $\ell_{\mathrm{n}}-$norm loss and high IoU, so IoU-based loss functions represented by Io$\mathrm{U}^{+}$1, Generalized IoU(GIoU), Distance IoU(DIoU) and Complete IoU(CIoU) appeared. Based on the simulation experiment of bounding box regression, we point out that the existing IoU-based loss functions have the problems of slow convergence speed, large training error, and poor training result. From the specific bounding box regression process, the causes of the problems are explored, and a new loss function LIoU combining smooth $\ell_{1}-$norm loss and IoU is proposed. The experimental results on the two-stage object detectors Faster R-CNN, Mask R-CNN and the single-stage object detector RetinaNet consistently show that LIoU gets rid of the problems of the existing IoU-based loss functions and can further improve the performance of the object detectors. The source code and model training weights are open source on https://github.com/iamRyanChiaLIoU.1IoU+ refers to the IoU-based loss function used in UnitBox [1].