基于三维立方体损失的点云目标检测算法-《集美大学学报（自然科学版）》

文章信息/Info

Author(s):: GE Xuyang1; HUANG Shangfeng2; CAI Guorong2; (1.School of Science,Jimei University,Xiamen 361021,China；2.College of Computer Engineering，Jimei University，Xiamen 361021，China)

摘要:: 常用的优化回归边界框参数的损失并不等价于最大化IoU指标，并且IoU作为回归损失，在边界框不重叠的情况下进行优化是不可行的。为了解决这个问题，在IoU基础上加入外接框的计算，将所得结果(3D_CGIoU)作为损失纳入到目前主流的三维目标检测框架中，并在数据集KITTI和ScanNet上进行实验。实验结果显示检测的平均精度得到了提升，表明该方法是有效的。

Abstract:: Optimizing the commonly used losses for regressing the parameters of a bounding box is not equivalent to maximizing the IoU index.And IoU as a regression loss,it is not feasible to optimize when boundary boxes do not overlap.In order to solve the above problems,the calculation of the circumscribed box is added to the basis of IoU(3D_CGIoU) as a new metric and new loss.Incorporating 3D_CGIoU as a loss into the stateoftheart 3D object detection framework,and experiments were conducted on the KITTI dataset and the ScanNet dataset respectively.The improvement of the average accuracy (AP) of the detection shows the effectiveness of the method.