Overview | Tasks

Using the two datasets we provide in Argoverse, we invite researchers to explore the utility of our rich maps with two common autonomous driving tasks: 3D tracking and motion forecasting.

Coming soon! We will maintain a leaderboard for users to benchmark the performance of their 3D tracking and motion forecasting methods, and see how they stack up against other submissions.


3D Tracking Task

The goal of the 3D tracking task is to annotate and track objects in 15-30 second log segments, using 17 object classes. To do this, users leverage the 113 sequences in the Argoverse 3D Tracking dataset.

The 89 training and validation sets include 3D cuboid annotations. The test sequences are limited to sensor data only.

training sequences

65

validation sequences

24

test sequences

24

A log segment with amodal 3D cuboid annotations, generated from data collected in Miami.

For more information on the 3D Tracking Task, see our tutorial.


Motion Forecasting Task

The goal of the motion forecasting task is to predict the location of a tracked object 3 seconds into the future, given an initial 2-second observation. To do this, users leverage the 327,793 sequences in the Argoverse Motion Forecasting dataset.

The training, validation, and test sequences are taken from different areas of Pittsburgh and Miami so that there is no geographical overlap. Each sequence includes one interesting tracked object, which we label as the “agent.” Agents are objects that follow more complex trajectories, such as changing lanes, navigating intersections, and turning.

training sequences

208,272

validation sequences

40,127

test sequences

79,391

Each training and validation sequence is 5 seconds long. The test sequences are 2 seconds long. In these, users are given the first 2 seconds (20 frames) and tasked with predicting the coordinates the agent will travel to within the next 3 seconds (30 frames) of the full 5-second segment (50 frames).

To explore this task ourselves, we ran a few baselines — one of which you can view in the qualitative results below. The orange trajectory represents the motion of the observed agent over the initial 2 seconds, the green trajectory(s) represents our top-k forecasted trajectory, and the red trajectory represents the ground truth.

For more information on the Motion Forecasting Task, see our tutorial.