Of all the video games you’d show somebody in driver’s ed, Grand Theft Auto may not be at the top of that list. But a team at Intel Labs and Darmstadt University in Germany has found that using video games gives an unparalleled level of accuracy when identifying objects.
The team, which published its findings in this paper noticed that the game provided an accurate simulation of real-world driving scenarios. This data can be used by self-driving cars in the real world to drive around and navigate safely.
Self-driving cars use object identification data to help “learn” how to identify objects like pedestrians, lampposts and walls when driving on the street. Normally, car makers create this data from recorded video from a car’s dashboard. They go through and identify objects manually, with the system using machine learning to eventually build up a broader idea of what each object looks like.
Using Grand Theft Auto, though, the team was able to automate this process far more effectively. The team could record similar videos in-game, but was able to identify assets faster that represented those same street objects. The photorealistic virtual world means the objects identified give the system the same accurate ideas of what real-world objects will look like.
The computer is able to automatically identify objects in mere seconds, a process that normally takes nearly two hours per image with recorded video. Here is the process in action:
“With artificial environments, we can effortlessly gather precisely annotated data at a larger scale with a considerable amount of variation in lighting and climate settings,” Alireza Shafaei, a Ph.D. student at the University of British Columbia, told MIT Technology Review.
Shafaei published his research in a paper that details how video games can train computers to help see the world. “We showed that this synthetic data is almost as good, or sometimes even better, than using real data for training,” he said.
Self-driving cars use a large amount of data, and techniques like these will be vital to keep on top of things. AT&T has begun trialing a new 5G cellular network, designed with self-driving cars in mind, that can prioritize mission-critical data to avoid driverless cars suffering from latency. All this data comes at a cost, though, as researchers have warned that cars could be susceptible to hacking. Driverless vehicles are opening up new possibilities for big data sets, but the question of how to handle it all will be the top priority.