The virtual training platform for computer vision

With the development of deep learning, computer vision has made great progress. Computer vision training based on deep learning requires a large number of data sets. However, manual obtaining of relevant data sets is costly, and some special samples are not easy to obtain. Therefore, in order to solve the lack of data sets, a virtual training platform is designed. The platform is able to render three-dimensional scene simulation by using Unity 3D, and automatically generates visual training data sets through corresponding script file. Through the interactive interface of the platform, users can select and adjust the model, scene, light and other variables independently according to different needs, and quickly generate multi-angle pictures and corresponding annotation information. Compared with the previous computer vision training data sets obtained by manual obtain of image information and label, it has the advantages of fast speed, low cost and large scale. At the same time, use the object detection network to evaluate the data generated by the platform and the data obtained by the previous way. It can be observed that the virtual data set can achieve certain effects. Therefore, the data sets generated by this platform can replace the traditional ones to a certain extent.