Nexar has released a dataset that it says is the world’s largest photo set featuring geographically diverse images for automotive tech development, for an open competition. There are 55,000 tagged photos in the set, taking from over 80 countries, in a variety of lighting and weather conditions. Each of the photos is taken from street level, using Nexar’s community-based V2V dashcam app for iOS and Android, and the goal of the release is to help drive the development of autonomous driving perception models that can handle a wide range of weather, road and country variety.
The release of Nexar’s image set, which it calls NEXET, is part of a challenge issued by the company to researchers to create a perception system for self-driving cars that’s able to work in a range of different settings, across geographical borders, while delivering consistent performance in all cases.
Nexar says that their goal is to address a significant gap in a lot of current research, which uses imagery for training that comes from either very circumscribed real world areas, or from simulations or lab-based environments. Any software developer knows that there are issues you only come across when dealing with real-world conditions, and that’s definitely true for training autonomous driving systems, which still face a huge hurdle in terms of addressing edge cases. With an iPhone app, outlier use cases have relatively low stakes; with driving, they could mean the difference between life and death.
Nexar’s whole goal is to build an Advanced Driver Assistance System that combines data from multiple streams via consumer devices around the world, and its competition is designed to help it further its own efforts. But the ultimate value to the industry is also apparent, and it’s rare to come across this size and type of dataset in the wild.