Mapillary Traffic Sign Recognition Dataset will Augment Traffic Sign Perception in Autonomous Vehicles

The Mapillary traffic sign recognition dataset is a first step toward teaching autonomous vehicles to perceive traffic signs using low-cost cameras.

FREMONT, CA: Research conducted earlier this year has revealed the possibility of leveraging cheaper cameras to aid autonomous vehicles in understanding their surroundings. The successful implementation of this technology would cut down the cost of autonomous vehicles by thousands of dollars. To achieve this result, Mapillary recently launched a traffic sign recognition dataset to teach autonomous vehicles to perceive traffic signs. The new dataset is the world’s largest and most diverse dataset publicly available today.

Mapillary is a street imagery platform which automates and scales mapping using computer vision. The Mapillary traffic sign dataset comprises over 100,000 images from around the globe. All the images have variability in factors ranging from time, weather, camera sensors, and viewpoints. Although traffic sign recognition technology is not uncommon, it has not been developed and enhanced to such an extent.

The data used by car manufacturers when training their algorithms often have low variability. When it comes to training data, increased diversity yields better results. Mapillary has chased a collaborative approach when developing the new product, using 570 million images uploaded by people and companies all across the world. Out of these, 100,000 images were selected for the traffic sign dataset. Over 300 different traffic sign classes are annotated and verified, which has enabled the labeling of over 320,000 images. 52,000 of the images have been verified and annotated by humans, whereas the rest have been annotated partially via the Mapillary’s computer vision technology.

The images were collected from regions including North and South America, Europe, Africa, and Asia, with high variability in weather conditions ranging from sunny to hazy and time ranging from dawn to dusk. A broad range of camera sensors of varying focal length, aspect ratio, and camera noise were leveraged to capture images from different viewpoints.

Mapillary’s dataset tackles a prominent obstacle of traffic sign recognition, which is a giant step toward camera-based solutions. The unique strength of the Mapillary traffic sign dataset lies in the wide variety of data used in its development. The process of collecting images from diverse locations from all over the world would prove too challenging for a single organization. With its solution, Mapillary aims to explore the potential of using only cameras to enable autonomous driving.