What do you need to teach driverless cars to drive? Masses of data, especially on traffic signs. And instead of leaving each automotive company to gather their own small scale data sets, the Swedish mapping company Mapillary thinks it can speed up the process.
The company is now releasing a new set of data of over 100,000 analysed images of traffic signs, which can be accessed for a licence fee. Mapillary, with about 570 million images on its platform, hopes that with the new data set specifically set on traffic signs, will increase the interest from the broader automotive industry.
“It’s like with us when we are taking our drivers licence, we need to train to understand all the traffic signs. That is the same thing that driverless cars need to go through now,” Emil Dautovic says.
Dautovic is the vice president of Automotive at Mapillary, the platform that started off as a collaborative service where people could use photos taken by a vast community of volunteer photographers and companies to edit and fix maps.
Recently the company has been getting a lot of interest from the automotive industry.
“Out of the five-year history of the company, it really started 18 months ago when the automotive segment took an interest,” Mapillary’s founder Jan-Erik Solem told Sifted earlier this year.
Same license – same prices for everyone. We don’t take sides, any company can use it.
“Imagery on the platform is available on an open data licence, on the side of that there is a standard commercial licence, same license – same prices for everyone. We don’t take sides, any company can use it. We monetise through commercial licenses of the imagery or the data we extract from the imagery,” Solem said.
When teaching driverless cars to drive, carmakers often use LiDAR, a way of measuring the surroundings with laser beams.
If you want self-driving technology to reach cars without costing a fortune, then a camera system is the way
But analysed photographs could be a better — and cheaper — way of doing this, according to a new paper by researchers at Cornell University. It has piqued the interest of Elon Musk among others and is good news for Mapillary, which can provide exactly this kind of data.
“Although LiDAR technology is good, it’s very expensive,” Dautovic says. “If you want self-driving technology to reach cars without costing a fortune, then a camera system is the way. Almost all cars today already have cameras or are starting to get them, even in cheaper cars.”
Not all of Mapillary’s millions of images are useful for the car industry, but the data set is diverse and from places all over the world, according to Dautovic. Other datasets are usually not as broad, which makes it more difficult for the algorithm to be correct when deciding what is in the picture.
“For example, San Francisco Bay Area is probably the world’s best-mapped area in the world. But driverless cars need to recognize traffic signs in other places as well,” says Dautovic.
“The strength in Mapillary really lies in the diversity of the input data. There are only a few other datasets on the market, and none of them has imagery from all over the world, simply because it would take too much effort for one, single player to get images from such a diverse set of locations on a global scale.”
San Francisco Bay Area is probably the world’s best-mapped area in the world. But cars need to recognize traffic signs in other places as well
For Mapillary’s community of thousands of people, from Denmark to Japan, however, it is easily do-able.
“So that’s why we believe the broadness in our data sets is good for algorithms,” Dautovic says.
So far Mapillary is working with car manufacturers such as Toyota, BMW and Daimler as well as the ride-hailing company Lyft.
By making its data available even more widely available by a set only focusing on traffic signs, it is hoping to get cars ready to drive by themselves even more quickly.
“Well, at least it helps in that process,” Dautovic says.