使用虚拟图片作为目标检测训练集

Application of Virtual Images to Object Detection Training Sets

对于基于深度学习的目标检测算法来说,想要获得高准确率,除了算法本身,还需要数以十万计的高质量图片作为训练集以及大批量的数据集标注。由于安全帽检测领域没有合适的大规模训练集,且除了网络上能够收集到的部分图片外,只能通过摄像头获取相关图片,但这种单一场景又导致模型过拟合而降低了泛化能力。为此研究一个能够通过人工生成的与真实场景相近的虚拟场景并进行自动标注,为缺少大规模训练集的检测模型提供一定的帮助的同时减少了手工标注工作。使用Unity3D建立了多个虚拟场景,并在场景内设置一些自由行动的佩戴、未佩戴安全帽的人员来获得相关图片和标注,使用Yolov3作为统一的检测程序进行检测,实验结果表明,以Unity3D为基础建立的虚拟图片能够帮助训练集较少的模型获得更高的检测精度。

In an object detection algorithm based on deep learning,to achieve high accuracy,in addition to the algorithm itself,hundreds of thousands of high-quality images are required in training sets and large data sets are annotated with an analyzed object image.Since there is no suitable large-scale training set in the field of helmet detection,and only relevant images can be obtained via camera in addition to the thousands of images that can be collected on the web,this single scenario leads to overfitting of the model and reduces the generalization capability.Therefore a virtual scenario is traind to generate similar to real one and automatically annotate,which provides some help to the detection model that lacks a large training set and reducing manual annotation work.Unity3D is adopted to build several virtual scenarios and set up some wearing helmeted or non-helmeted people within the scenarios to get relevant images and annotations.Experimental results on detection performing by using Yolov3 as a unified detection procedure show that the virtual images built on Unity3D can help to obtain higher detection accuracy for model with a smaller training set.