CVPR 2022: How Argo AI is Advancing Object Detection and Perception for Self-Driving Cars

A point cloud image shows a city street through the view of an Argo Lidar sensor, with other nearby traffic rendered in teal and further objects rendered in blue, pink, and purple

The annual IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) begins next week in New Orleans, Louisiana — the largest CVPR ever and its first in-person event in nearly three years. CVPR is an important event for computer vision and machine learning researchers, bringing them together from around the world to review, discuss, and build upon their latest breakthroughs in autonomous vehicles and other industries impacted by artificial intelligence.

Leading autonomous vehicle technology company Argo AI is attending and sponsoring CVPR 2022 to share the company’s latest research from its principal scientists and their colleagues and discuss the advancements its growing engineering team is making in Lidar development, perception, object tracking, detection, and other computer vision topics. Leaders from Argo AI will also be discussing new developments in 3D object detection and motion forecasting based on new datasets and high-definition maps that the company released publicly to support academic research earlier this year.

Argo’s presence at CVPR 2022 includes:

  • Publication of 7 different research papers relevant to perception and forecasting for self-driving vehicles either directly from Argo engineers or with our university partners.
  • Presentations at CVPR’s Workshop on Autonomous Driving on the advancements being made in perception based on the Argoverse 2 public dataset, as well as the winners of the Argoverse 2022 Summer Challenges
  • Running and sponsoring the MOT Challenge event on how synthetic data can enable multi-object tracking
  • Through our partnership with Carnegie Mellon University (CMU), running and sponsoring the Visual Perception and Learning in an Open World workshop that explores visual understanding for open-world, continually evolving data streams
  • A booth on the CVPR Expo floor, where the company will display an Argo Lidar equipped self-driving car, and share how Argo AI is building computer vision systems that support object forecasting, detection and perception for its self-driving vehicle fleet

Come stop by booth #1407 on the CVPR Expo floor to learn more about Argo AI’s career opportunities and the company’s work bringing autonomous vehicles to scale in the U.S. and Germany. See Argo’s open positions here.

The CMU Argo AI Center for Autonomous Vehicle Research on Forecasting from Lidar via Future Object Detection

Object detection and forecasting are fundamental components of perception for self-driving cars. However, they are largely studied in isolation.

The CMU Argo AI Center for Autonomous Vehicle Research’s paper, Forecasting from LiDAR via Future Object Detection, presents a new method for detecting moving objects and forecasting their future trajectories entirely from lidar data. This end-to-end approach — in which an algorithm takes in just a second of lidar data from an autonomous vehicle and produces estimates of future behaviors of nearby objects — is among the first-of-its-kind. The method generates multiple future paths for each object in lidar, and ranks them in order of most to least likely. The research shows how lidar processing, which is commonly used in the autonomous vehicle industry for real-time detection of moving objects, can be expanded into new use cases such as motion forecasting, providing further efficiencies, utility, and accuracy for self-driving systems.

The CMU Argo AI Center for Autonomous Vehicle Research has 6 other papers publishing at CVPR on perception and object detection. Read more on Ground Truth.

New Innovations in 3D Object Detection and Motion Forecasting from Argoverse Data

Diverse perspectives are needed in building and scaling autonomous driving systems. As self-driving cars begin to scale to more cities across the globe, Argo AI is committed to identifying and filling critical gaps in the public data available to academic researchers.

Earlier this year, Argo AI announced the second release of Argoverse data (Argoverse 2), which includes one of the largest collections of open-source autonomous driving data and high-definition maps from six U.S. cities: Austin, Detroit, Miami, Pittsburgh, Palo Alto, and Washington, D.C. This dataset enables advanced research into algorithms that can detect out-of-date HD maps and new methods of machine learning using unlabeled point cloud data.

To encourage advanced research into core areas of autonomy, Argo AI launched two competitions that utilize the Argoverse datasets to challenge participants from around the world to innovate in 3D object detection and motion forecasting. At CVPR’s Workshop on Autonomous Driving, James Hays, Principal Scientist at Argo AI, will reveal the winners from these two competitions and highlight the innovations the winners developed based on the new Argoverse data.

Argo Lidar Advancements for Widespread Autonomous Rideshare & Deliveries

At Argo, we believe Lidar technology is essential for building self-driving cars that can safely operate on city streets and highways amongst other vehicles, pedestrians, bicyclists and other road users. Argo designed and developed its own long-range lidar in house, which has given the company a competitive edge in deploying self-driving cars in eight cities across the globe.

Argo Lidar gives Argo AI’s vehicles 360-degrees visibility day and night to drive on busy city streets, suburban neighborhoods, and at highway speeds. This technology has long-distance sensing range capabilities of over 400 meters, and can detect the smallest particle of light to sense objects with low reflectivity. People and lidar see light differently, and Argo Lidar can see dark objects at far distances that are not always visible to the human eye.

Argoverse 2 has one of the largest lidar datasets in the autonomous driving industry. Releasing lidar datasets provides academics, researchers and engineers with the information needed to help advance key aspects of safe and efficient autonomous driving, such as point cloud forecasting or self-supervised learning (a method of machine learning with unlabeled data).