Getting good labeled datasets is a challenge in any AI-training environment, and self-driving vehicles are no exception. But with all the self-driving car accidents lately, it's worth looking at taking all the training risks in a virtual world instead. Scientists in China have developed just such a solution, creating a VR version of an area of Beijing in the Unity 3D engine, which is used to develop many video game titles. The system, known as ParallelEye, then captures images with a virtual camera mounted on a virtual car driving through virtual city streets in a variety of virtual weather conditions (another challenge to replicate on command in the real world). These artificially-generated datasets were then combined with real-world ones, and the accuracy of the models improved. It's a work in progress, for sure, but synthetic training is an encouraging option in a time when so much of the world's data is generated and collected by a handful of tech giants.
sign up for our newsletter