Overview | Argoverse 1

Terms of Use

We created Argoverse to support advancements in 3D tracking, motion forecasting, and other perception tasks for self-driving vehicles. We offer it free of charge under a creative commons share-alike license. Please visit our Terms of Use for details on licenses and all applicable terms and conditions.

Citation

In June 2019, we released the Argoverse 1 datasets to coincide with the appearance of our publication, Argoverse: 3D Tracking with Forecasting and Rich Maps, in CVPR 2019. When referencing this publication or any of the materials we provide, please use the following citation:

@INPROCEEDINGS { Argoverse,
  author = {Ming-Fang Chang and John W Lambert and Patsorn Sangkloy and Jagjeet Singh
       and Slawomir Bak and Andrew Hartnett and De Wang and Peter Carr
       and Simon Lucey and Deva Ramanan and James Hays},
  title = {Argoverse: 3D Tracking and Forecasting with Rich Maps},
  booktitle = {Conference on Computer Vision and Pattern Recognition (CVPR)},
  year = {2019}
}

Where was the data collected?

The data in Argoverse 1 comes from a subset of the area in which Argo AI’s self-driving test vehicles are operating in Miami and Pittsburgh — two US cities with distinct urban driving challenges and local driving habits. We include recordings of our sensor data, or "log segments," across different seasons, weather conditions, and times of day to provide a broad range of real-world driving scenarios.

Total lane coverage: 204 linear kilometers in Miami and 86 linear kilometers in Pittsburgh.

How was the data collected?

We collected all of our data using a fleet of identical Ford Fusion Hybrids, fully integrated with Argo AI self-driving technology. We include data from two lidar sensors, seven ring cameras and two front-facing stereo cameras. All sensors are roof-mounted:

Argoverse 1 Maps

Our maps contain rich geometric and semantic metadata for better 3D scene understanding. From ground height to the distance remaining until the next intersection, our maps enable researchers to explore the potential of HD maps in robotic perception.

There are three distinct components that set our maps apart:

Vector Map: Lane-Level Geometry

Our semantic vector map conveys useful lane-level detail, such as lane centerlines, traffic direction, and intersection annotations. Through these features and more, users can explore the many ways traffic flows through city streets and complicated intersections in our test areas, and access a comprehensive picture of what comes before and after each scene.

Rasterized Map: Drivable Area

Our maps include binary drivable area labels at one-meter grid resolution. A drivable area is an area in which it is possible for a vehicle to drive (though not necessarily legally). Our track annotations, outlined in 3D Tracking, extend to five meters beyond the drivable area. We call this larger area our "region of interest."

Rasterized Map: Ground Height

Our maps include real-valued ground height at one-meter resolution. With our map tools, users can remove lidar returns on uneven ground surfaces for easier object detection.

Argoverse 1 Map API

Our Map API enables users to more easily leverage our rich data and even take advantage of open source mapping tools built to handle OpenStreetMap formats.

We provide this API in Python and include a handful of useful functions, such as:

Argoverse 1 3D Tracking

A dataset to train and validate 3D tracking models

Argoverse 1 3D Tracking is a collection of 113 log segments with 3D object tracking annotations. These log segments, which we call “sequences,” vary in length from 15 to 30 seconds and collectively contain a total of 11,052 tracks.

Each sequence in our training and validation sets includes annotations for all objects within 5 meters of what we identify as the “drivable area” — the area in which it is possible for a vehicle to drive.

What makes this dataset stand out?

Users can build algorithms that take advantage of the detailed information in Argoverse's HD Maps. For example, an algorithm can use the map to perform ground removal on lidar returns or to constrain vehicle orientation based on lane direction.

Data Annotation

Argoverse 1 contains amodal 3D bounding cuboids on all objects of interest on or near the drivable area. By “amodal” we mean that the 3D extent of each cuboid represents the spatial extent of the object in 3D space — and not simply the extent of observed pixels or observed lidar returns, which is smaller for occluded objects and ambiguous for objects seen from only one face.

Our amodal annotations are automatically generated by fitting cuboids to each object’s lidar returns observed throughout an entire tracking sequence. If the full spatial extent of an object is ambiguous in one frame, information from previous or later frames can be used to constrain the shape. The size of amodal cuboids is fixed over time. A few objects in the dataset dynamically change size (e.g. a car opening a door) and cause imperfect amodal cuboid fit.

To create amodal cuboids, we identify the points that belong to each object at every timestep. This information, as well as the orientation of each object, come from human annotators.

We provide ground truth labels for 15 object classes. Two of these classes include static and dynamic objects that lie outside of the key categories we defined, and are called ON_ROAD_OBSTACLE and OTHER_MOVER. The distribution of these object classes across all of the annotated objects in Argoverse 3D Tracking looks like this:

Argoverse 1 Stereo Dataset

A dataset to train and validate stereo depth estimation algorithms

Argoverse 1 Stereo consists of rectified stereo images and ground truth disparity maps for 74 out of the 113 Argoverse 1 3D Tracking Sequences. The stereo images are (2056 x 2464 px) and sampled at 5 Hz. The dataset contains a total of 6,624 stereo pairs with ground truth depth, although we withhold the ground truth depth for the 15-sequence test set.

To create the ground truth depth, we use our recent scene flow method to accumulate lidar points from 11 frame sliding windows. We also manually refine the calibration and discard sequences with too many spurious lidar returns.

What makes this dataset stand out?

Inspired by the KITTI stereo benchmark, our stereo dataset has 16 times as many frames and features more diverse driving scenarios. We also accumulate depth differently by relying on non-rigid scene flow as opposed to vehicle CAD model fitting, which KITTI uses.

The Argoverse lidar range is up to 200 meters, roughly twice the range of the sensor used in the KITTI dataset, and our lidar produces a point cloud with greater density at each sweep. Therefore, our stereo dataset captures more scene details at the near and far range, crucial for autonomous vehicles.

Argoverse 1 Motion Forecasting Dataset

A dataset to train and validate motion forecasting models