基于YOLOv5s模型的边界框回归损失函数研究

Research on bounding box regression loss function based on YOLOv5s model

针对车辆检测中边界框回归损失函数与检测目标尺度不匹配导致的误检、漏检以及精度较低等问题,基于YOLOv5s模型对4种有代表性的边界框回归损失函数进行对比实验,并在UA-DETRA、VisDrone2019、KITTI数据集上进行验证,利用漏检率、误检率、准确率、召回率、mAP@0.5、迭代过程的边界框损失值以及目标检测结果对其适用场景进行分析研究。结果显示:CIoU整体性能最差;SIoU在KITTI数据集上整体性能最优,准确率最高,达到了94.5%,漏检率降到了1.2%,适用于中尺度目标检测任务;Focal-EIoU在VisDrone2019数据集中各项指标远优于其他损失函数,召回率和mAP@0.5指标相较于CIoU分别提高了1.6%和1.8%,误检率降低了6.9%,且迭代过程损失值远低于其他损失函数,适用于小尺度目标检测任务;WIoU在UA-DETRA数据集整体性能最优,漏检率、召回率以及mAP@0.5指标优于其他损失函数,适用于大尺度目标检测任务。此研究为目标检测任务的边界框回归损失函数的选择提供了重要的基础。

In view of the false detection,missed detection and low precision caused by the mismatch between the bounding box regression loss function and the detection object scale in vehicle detection,four representative bounding box regression loss functions are contrasted based on the YOLOv5s model,and verified on the datasets of UA⁃DETRA,VisDrone2019 and KITTI.The missed detection rate,false detection rate,precision,recall rate,mAP@0.5,the bounding box loss value of the iterative process and the object detection results are used to analyze and study the applicable scenarios.The results show that the overall performance of CIoU is the worst,SIoU has the best overall performance on the dataset KITTI,with the highest precision of 94.5%,and its missed detection rate is reduced to 1.2%,which is suitable for the detection tasks of the objects with medium scale.Focal⁃EIoU is far superior to the other loss functions on the data set VisDrone2019.In comparison with CIoU,its recall rate and mAP@0.5 indicators are improved by 1.6%and 1.8%,and its false detection rate is reduced by 6.9%,and its loss value of the iterative process is much lower than the other loss functions,that is,it is suitable for the detection tasks of the objects with small⁃scale.WIoU has the best overall performance on the dataset UA⁃DETRA,and its missed detection rate,recall rate and mAP@0.5 are better than those of the other loss functions,which is suitable for the detection tasks of the objects with large⁃scale.This study provides an important basis for the selection of bounding box regression loss function for object detection tasks.