摘要
为了解决多点表示方法在鱼眼图像人体检测过程中标注成本高的问题,本文提出了基于FastSAM的多点标注算法,将鱼眼图像数据集本身的矩形框标注作为提示引导框,与原图一起作为输入,通过FastSAM获取较为准确的目标分割标注,根据分割标注与矩形框的IoU评估分割信息的准确性,对于不准确的个体进一步补全纠错。针对多点表示无法处理中心点不在目标内的问题,提出了基于凸包的多点表示回归策略,直接通过分割信息获取多点表示标注信息,同时设计了相对应的标签分配机制和损失函数。本文的方法可以节省大量的人工成本,同时通过实验验证该算法具有可行性。
The multipoint representation method has significant advantages in the field of people detection in fisheye images.However,its annotation process is time-consuming and labor-intensive.Therefore,this paper proposes a multipoint annotation algorithm based on FastSAM.First,the rectangular bounding boxes annotated on the fisheye image datasets are used as prompt boxes,which are input into FastSAM along with the original image to obtain more accurate object segmentation annotations.The accuracy of the segmentation information is evaluated according to the IoU between the segmentation annotation and the prompt box.For inaccurate annotations,further manual screening and correction are performed.To address the issue of the multipoint representation's inability to handle cases in which the center point is not inside the target,we propose a convex hull-based multipoint representation regression strategy.This strategy can directly obtain multipoint representation annotations through segmentation information,and a corresponding label assignment mechanism and loss function are designed.The method in this paper can save a lot of labor costs,and the feasibility of the algorithm is verified through experiments.
作者
乔人杰
蔡成涛
QIAO Renjie;CAI Chengtao(College of Intelligent Systems Science and Engineering,Harbin Engineering University,Harbin 150001,China;Heilongjiang Provincial Key Laboratory of Environment Intelligent Perception,Harbin 150001,China;Key Laboratory of Intelligent Technology and Application of Marine Equipment of the Ministry of Education(Harbin Engineering University),Harbin 150001,China)
出处
《哈尔滨工程大学学报》
EI
CAS
CSCD
北大核心
2024年第8期1427-1433,共7页
Journal of Harbin Engineering University
基金
黑龙江省自然科学基金重点项目(ZD2022F001).
关键词
多点标注
鱼眼图像
人体检测
FastSAM
凸包
提示矩形框
标签分配
损失函数
multipoint annotation
fisheye image
people detection
FastSAM
convex hull
prompt rectangular box
label assignment
loss function
