摘要
边界框回归(BBR)已广泛应用于目标检测和实例分割,这是目标定位的一个重要步骤,但仍存在收敛缓慢和回归不准确的问题.本文研究发现大多数现有的边界框回归损失函数在预测框与标注框具有相同的纵横比,但宽度和高度值不同时损失函数值无法收敛.为了解决这个问题,本文充分挖掘矩形的几何特征,提出了一种最小点距离的边界框相似度度量,它包含了现有主流边界框回归度量的相关因素,即重叠或非重叠面积、中心点距离、宽度和高度的偏差,同时简化了计算过程.在此基础上,本文提出了一个最小点距离的边界框回归损失函数,称为.实验结果表明,损失函数应用于最先进的实例分割(例如YOLACT)和目标检测(例如YOLOv7)模型训练PASCAL VOC、MS COCO和IIIT5k,其性能优于现有损失函数,模型回归效率和精度得到有效提升.
Bounding box regression(BBR)has been widely used in object detection and instance segmentation,which is an important step in object localization,but there are still problems of slow convergence and inaccurate regression.In this paper,it is found that most existing bounding box regression loss functions have the same aspect ratio as the groundtruth box in the predicted box,but the loss function values cannot converge when the width and height values are different.In order to solve this problem,this paper fully explores the geometric characteristics of rectangles and proposes a bounding box similarity metric with minimum point distance,which contains the relevant factors of the existing mainstream bounding box regression metric,namely overlapping or non-overlapping area,center point distance,width and height deviation,and simplifies the calculation process.On this basis,this paper proposes a bounding box regression loss function for the minimum point distance,called.Experimental results show that the loss function is applied to training PASCAL VOC,MS COCO and IIIT5k on the most advanced instance segmentation(e.g.,YOLACT)and object detection(e.g.,YOLOv7)models,and its performance is better than the existing loss function,and the model regression efficiency and accuracy are effectively improved.
作者
麻斯亮
许勇
MA Siliang;XU Yong(School of Computer Science and Engineering,South China University of Technology,Guangzhou 510006,China;Pengcheng Laboratory,Shenzhen 518000,China)
出处
《小型微型计算机系统》
CSCD
北大核心
2024年第11期2695-2701,共7页
Journal of Chinese Computer Systems
基金
国家自然科学基金项目(62072188)资助.
关键词
目标检测
实例分割
边界框回归
损失函数
最小点距离
object detection
instance segmentation
bounding box regression
loss function
minimum point distance
