Argoverse Gives Researchers Access to New Datasets for Autonomous Vehicles

Argoverse is the first public autonomous vehicle dataset to include high-definition (HD) maps.

Developing autonomous vehicles has long been a hot topic in pop culture and the tech community, but the material that’s needed to further academic research — data from autonomous vehicle sensors and other telemetry --  is usually kept under lock and key. Researchers and engineers at Argo AI and Georgia Institute of Technology recently challenged that by releasing Argoverse, the first public autonomous vehicle dataset to include high-definition (HD) maps.

Argoverse’s HD maps contain accurate details to within a few centimeters. These maps help autonomous vehicles better understand the rules of the road through geometric and semantic metadata such as where a driver should stop for an intersection, what the travel direction is for a particular lane, and what turns are available in each lane, if any. And when it comes to research, the maps can be used to develop more accurate forecasting models by painting a broader and more accurate picture of road infrastructure and traffic flow.

In addition to HD maps, Argoverse contains two datasets that allow researchers to train and benchmark 3D object tracking and forecasting methods. When applied to the autonomy stack, these methods play a critical role in enabling autonomous vehicles to identify objects on the road -- such as other cars, bicyclists and pedestrians -- track them over time, and forecast their behavior seconds into the future.

Having access to high-quality maps and curated data collections is critical to furthering autonomous vehicle research. Argoverse comes at a time when many experts and academics in the field can benefit from materials that take tremendous resources and capital to produce. Building a single autonomous vehicle could cost upwards of a few hundred thousand dollars, and that has to be done before putting the tools in place to build a map.

“It could be seen as a competitive disadvantage for a company to release data like this, but over the past few years the industry has started to realize the benefits of engaging the academic community,” said James Hays, an associate professor in the School of Interactive Computing at Georgia Tech and a principal scientist at Argo AI. “Creating autonomous vehicles is a big challenge that combines so many aspects of technology. By putting out this dataset, Argo is providing material for others to discover clever ways to improve self-driving capabilities.”

Argo AI said in a statement, “For our team at Argo, releasing this data collection is about giving academic communities access to the materials they need. We’re excited to not only support cutting edge developments in computer vision and machine learning but also to support the next generation of engineers and roboticists who are preparing for jobs at self-driving technology companies, Argo AI included.”

Inspired by the KITTI dataset, Argoverse includes one dataset with 3D tracking annotations for 113 scenes and one dataset with 327,793 interesting vehicle trajectories extracted from over 1,000 driving hours. The Argoverse data collection also includes an API to connect sensor data with the HD map representation.

The HD map has three layers: One layer encodes the ground height at any location, while another layer indicates the drivable area. The most complex layer encodes the geometry and connectivity of individual lanes.

The data was collected in Miami, Fla. and Pittsburgh, Penn. – covering 180 linear miles of the two distinct urban cities that each possess unique local driving habits and challenges.

Two of the paper’s lead authors, John Lambert and Patsorn Sangkloy, are Ph.D. students in Hays’s lab at Georgia Tech and presented the paper Argoverse: 3d Tracking and Forecasting with Rich Maps at the 2019 Computer Vision and Pattern Recognition (CVPR) conference.

Contact: 

Allie McFadden

Communications Officer

allie.mcfadden@cc.gatech.edu