摘要
不同目标检测模型的标注格式是不一样的。比较评估多种模型时,需要对标注文件进行数据打包,获得对应模型的标注数据格式。本文提出一种简便的数据打包方案,即利用通用的VOC格式作为转换桥梁,只需要少量编程,就可以转换为大部分主流检测模型的标注数据格式。首先创建对应的VOC文件夹结构;再根据图片名称遍历整个原标注文件,同时调整图片类型与尺寸,并根据调整后的图片尺寸比率转换标注框数据;获得对应的VOC格式标注文件。实验将交通标志检测数据集GTSDB的标注文件转换为YOLO格式,结果表明该数据打包方案能够获得准确的标注文件信息,方便对比目标检测模型的精度及速度等指标。
The annotation formats of different objection detection models are different.When comparing and evaluating different models,it is necessary to package the annotation files to obtain the annotation data formats of corresponding models.In this paper,a simple data packaging scheme is proposed,which uses the general VOC format as the transformation bridge,which can be converted into the annotation data format of most mainstream detection models with a small amount of programming.Firstly,the corresponding VOC folder structure is created.Then the whole original annotation file is traversed according to the image name,adjust the image type and size,and convert the bounding box data according to the adjusted image size ratio.Finally,the corresponding VOC format annotation file is obtained.In the experiment,the traffic sign detection dataset file of GTSDB is transformed into YOLO format.The experimental results showthat the data packaging scheme can obtain accurate annotation file information,which is convenient for comparing the accuracy and speed of objection detection model.
作者
张光建
朱婵
ZHANG Guangjian;ZHU Chan(Department of Fundamental Subjects,Sichuan College of Architectural Technology,Deyang Sichuan 618000,China)
出处
《智能计算机与应用》
2020年第7期220-223,共4页
Intelligent Computer and Applications
基金
德阳市重点科学技术研究项目(2017ZZ058)
四川建院科学技术研究项目(2017KJ11)
