Public datasets supported by detailed maps to train and test methods for perception and prediction in driving scenes.

The first self-driving datasets with...

...detailed maps

Argoverse maps include geometric and semantic metadata, such as lane boundaries and driveable area. These details make it possible to develop more accurate perception algorithms, which in turn will enable self-driving vehicles to safely navigate complex city streets.

...hundreds of thousands of interesting scenarios for motion forecasting

Motion forecasting, the task of predicting the location of tracked objects in the future, is an essential component of developing safe self-driving systems. To support advanced research into motion forecasting, Argoverse includes two motion forecasting datasets: the original Argoverse 1 Motion Forecasting Dataset and its successor, the Argoverse 2 Motion Forecasting Dataset, which contains longer sequences and richer object attributes.

...tens of thousands of lidar scenarios backed by HD maps

Recent advances in machine learning have shown that surprisingly powerful representations can be learned from unannotated data. The Argoverse 2 Lidar Dataset is one of the largest lidar datasets in the autonomous driving industry, with a staggering 6 million lidar frames and 20,000 scenarios. With this dataset, researchers can advance key aspects of safe and efficient autonomous driving, from point cloud forecasting to self-supervised learning.

...HD map change observations

High definition (HD) maps act as a blueprint for the many fixed objects autonomous vehicles encounter on the road, from stop signs to lane boundaries. As real-world environments change, it is critical that HD maps reflect those changes in order to continuously provide the most up-to-date information to the autonomous driving system. To support the development of novel methods for detecting out-of-date map regions, Argoverse 2 includes a map change dataset.

Inside Argoverse

The Argoverse 1 open-source data collection includes:

  • A 3D Tracking Dataset with 113 3D annotated scenes
  • A Motion Forecasting Dataset with 324,557 scenarios

The Argoverse 2 open-source data collection includes:

  • A Sensor Dataset with 1,000 3D annotated scenarios — each with lidar, ring camera, and stereo sensor data
  • A Lidar Dataset with 20,000 unlabeled scenarios suitable for self-supervised learning
  • A Motion Forecasting Dataset with 250,000 interesting driving scenarios with richer attributes than its predecessor, the Argoverse 1 Motion Forecasting Dataset
  • A Map Change Dataset with 1,000 scenarios, 200 of which depict scenes that changed since mapping

Dataset details