The discussion around self-driving cars, and whether and when they’ll actually arrive, is never-ending. Forecasts and predictions range from “they’re already here” to “they’ll never arrive”. Meanwhile, researchers worry about the unpredictability of pedestrians, and pedestrians worry about the vigilance of autonomous vehicles. At the core of each discussion is the data that’s needed to train the cars.
Teaching cars to see, understand what they’re seeing, and make a subsequent decision on how to behave is a monumental task. A human driver has a lifetime experience of gathering and analysing data from streets. We’re well familiar with pedestrians, cars, cyclists, and even animals like horses in traffic environments. We’re usually able to make sound assessments of radically different traffic situations, simply because we’ve had a lifetime training in assessing such different situations.
Machines face the same task, but their understanding is limited to the data they’ve been exposed to. That’s why they have to be fed and trained on gigantic masses of data to properly understand street scenarios that would be uncomplicated to human drivers. A diverse dataset is absolute key. If a car’s system is trained on, say, a dataset of images from just Europe or the US, that system will inevitably have a narrow understanding of street scenarios. Street scenes vary considerably depending on landscape, weather, time of day, not to mention the city and country you’re in. To top it all off, streets in even developed countries like the US change by up to 15% annually – so autonomous cars need to be trained on wildly different street scenes to be able to cope with the pace that streets are changing.
Expecting just one company or organization to gather such masses of data with the variety needed is very optimistic, not to say unrealistic. It would require enormous investments in time and money, and even then, it’d be impossible for anyone to be everywhere.
The good news is, people and organizations are already sharing street-level images from all over the world. Automotive players don’t need to send fleets of cars to neither Sweden nor India or Silicon Valley to gather data, because people in those countries and beyond are already sharing their street-level images with the world. Cities like Amsterdam, countries like Lithuania, and organizations like Departments of Transport across the US have opted to share millions of street-level images with the public through platforms like Mapillary, making the data available to anyone. As a result, companies developing autonomous systems are in a position to access data from across the world to speed up the training of their own systems. Getting access to such wide and varied data would simply not be possible if it had been kept proprietary.
Sharing data in the form of street-level imagery won’t solve every problem in autonomous driving. Pedestrians might still worry about cars’ vigilance, and researchers might still worry about pedestrians’ jaywalking. However, sharing data will allow autonomous systems to get smarter, faster. This will only ever help us speed up the race towards an autonomous future.
About the Author
Emil Dautovic is VP of Automotive at Mapillary, the street-level imagery platform that uses computer vision to fix the world’s maps. The Mapillary Vistas Dataset is the world’s most diverse street-level imagery dataset for training autonomous systems, and is free of charge for all research purposes.