Abstract
Bounding box regression is one of the important steps of object detection.However, rotation detectors often involve a more complicated loss based onSkewIoU which is unfriendly to gradient-based training. Most of the existingloss functions for rotated object detection calculate the difference betweentwo bounding boxes only focus on the deviation of area or each points distance(e.g., $\mathcal{L}_{Smooth-\ell 1}$, $\mathcal{L}_{RotatedIoU}$ and$\mathcal{L}_{PIoU}$). The calculation process of some loss functions isextremely complex (e.g. $\mathcal{L}_{KFIoU}$). In order to improve theefficiency and accuracy of bounding box regression for rotated objectdetection, we proposed a novel metric for arbitrary shapes comparison based onminimum points distance, which takes most of the factors from existing lossfunctions for rotated object detection into account, i.e., the overlap ornonoverlapping area, the central points distance and the rotation angle. Wealso proposed a loss function called $\mathcal{L}_{FPDIoU}$ based on fourpoints distance for accurate bounding box regression focusing on faster andhigh quality anchor boxes. In the experiments, $FPDIoU$ loss has been appliedto state-of-the-art rotated object detection (e.g., RTMDET, H2RBox) modelstraining with three popular benchmarks of rotated object detection includingDOTA, DIOR, HRSC2016 and two benchmarks of arbitrary orientation scene textdetection including ICDAR 2017 RRC-MLT and ICDAR 2019 RRC-MLT, which achievesbetter performance than existing loss functions.