Towards Global Traffic Sign Recognition

We are taking a big step towards recognizing traffic signs all over the world by adding support for more than 500 traffic signs globally, together with an appearance-based taxonomy for traffic signs. This is the beginning of our journey of recognizing every road sign in the world, no matter where it is.

At Mapillary, we are on a mission to extract meaningful data from images. We find and recognize objects in geotagged photos automatically with computer vision and deep neural networks. Traffic signs are the first detections we started to do, since they are one of the key map features for navigation. What is more, with the accelerating development of autonomous driving, the relevance of this cannot be underestimated.

With the release of the first traffic sign recognition two years ago, we were able to recognize traffic signs in a few European countries and the United States. Half a year later, we added traffic sign recognition for Brazil, Canada and Australia. Since then, we have been looking into extending traffic sign recognition globally. With images contributed from all over the world by our community, we are in possession of an extremely diverse dataset to train our algorithms on.

As an important milestone towards global traffic sign recognition, we are now extending the existing traffic sign recognition with more than 500 signs in over 60 countries such as China, South Africa, Japan, Mexico, Slovenia, etc.

Traffic sign examples Examples of added traffic signs in China, Mexico, Japan, Slovenia and USA

Appearance Groups

To enable traffic sign recognition, we use deep neural networks and we need to have data about the appearances of traffic signs across different countries together with their corresponding meanings. During the data collection process we observed that many countries use signs with a similar or even the same appearance. For instance, the tunnel warning signs in Romania and the UK look almost the same.

Tunnel warning signs
Tunnel warning sign in Romania (left) and the UK (right)

Because of these similarities, training country-specific neural networks is not necessary (as many networks are learning to classify similar signs). At the same time, treating these similar traffic signs as different traffic signs will also be very ineffective, as it requires the neural networks to learn the subtle differences between them.

To handle the traffic sign extension systematically and efficiently, we have therefore decided to merge traffic signs into appearence groups. Each appearance group represents a set of sign appearances with little variation between different countries. That way, we effectively reduce the number of traffic sign classes for the deep neural networks to recognize. As a result, we have successfully trained fast and accurate deep neural networks for hundreds of traffic sign classes.

Take a look at the complete list of supported traffic sign classes and their representative appearances on our developer documentation page. We have set up processing with the updated recognition pipeline for millions of photos in all the supported countries. In the coming weeks, developers will be able to access and use the updated signs through our API, e.g. for navigation apps or map editing. The OpenStreetMap community can easily access the data in iD editor as well. Here are some examples of the recognition on real images.

Chinese traffic signs
Chinese warning signs
Japanese traffic signs
Japanese signs
Highway traffic sign
Highway interchange signs
Polar bear traffic sign
Polar bear sign

Moving forward, we are aiming to extend the traffic sign recognition to more countries, adding further support for complementary signs and optical character recognition (OCR) for textual signs. We would love to hear your thoughts on this update so feel free to comment, tweet or email us.

/Yubin and the vision team

Continue the conversation