List of Datasets

Automotive

Argoverse 2 Sensor

Multimodal driving perception and 3D tracking

One thousand fifteen-second driving scenarios with nine cameras, paired LiDAR sweeps, tracked 3D boxes and scenario-local HD maps.

Metadata: Complete Metadata source documented Released 2022

Primary places to inspect, download, read about, or reproduce the dataset.

Quick facts

Vehicle type
Automotive
Environment
Outdoor
Data origin
Real-world
Sensor count
11
Ground truth
Available
Calibration
Parameters Only
Annotations
Available

Sensor overview

High-level sensor availability before the detailed sensor records below.

Cameras

9 reported

IMUs

Not reported

GNSS

Not reported

LiDAR

2 reported

Radar

Not reported

Magnetometers

Not reported

Barometers

Not reported

UWB

Not reported

Ground truth

Reference scope, method, rate, coverage, and provenance.

Scope Method Reference system Count Rate Coverage Reference provenance
External
Measured using equipment independent of the platform’s normal onboard sensor pipeline.
Derived
Computed primarily from the dataset’s own recorded sensors.
Hybrid
Combines independent reference equipment with onboard measurements.
6DoF Other Map-aligned ego pose Not reported Not reported Unknown Reference provenance
External
Measured using equipment independent of the platform’s normal onboard sensor pipeline.
Derived
Computed primarily from the dataset’s own recorded sensors.
Hybrid
Combines independent reference equipment with onboard measurements.
Derived
Reference notes
Official AV2 page states combined GPS-/sensor-based 6-DoF localization. The map-aligned estimate depends on recorded sensing; detailed fusion stages, valid intervals and independent errors remain unknown.

Calibration and synchronization

Reported calibration level, reproducibility signals, and supporting notes.

Level
Parameters Only
Processed parameters
Available
Raw calibration data
Not reported
Calibration targets
Not reported

Parameters only

Processed calibration parameters are reported, but raw calibration data is not reported.

Notes

Per-scenario camera intrinsics and sensor extrinsics are documented. Full raw calibration sessions and independently measured release-wide residual errors remain unknown.

Known or intentional limitations

Reported constraints and characteristics to check before using the dataset.

One thousand roughly fifteen-second scenarios are a curated perception product, distinct from the AV2 LiDAR and Forecasting releases. Nominal clip sums do not establish unique non-overlapping driving time.

Six ring cameras use landscape images and the front-centre ring camera is portrait; the forward stereo pair adds two optical streams. Two physical 32-beam LiDARs are aggregated into compensated sweeps, not one 64-beam device.

Map-aligned ego poses combine GPS- and sensor-based localization. The detailed fusion/validation and independent numerical accuracy are not established; raw GNSS/IMU inputs are not inferred.

Thirty native annotation classes differ from twenty-six evaluated classes. Track annotations concern objects near the drivable region and task split coverage; amodal cuboids can imperfectly fit objects changing shape.

Cameras

Camera 1

Model
Not reported
Setup
Mono
Count
6
Effective cameras
6
Modality
Rgb
Resolution
2048 x 1550
Rate
20 Hz
Shutter
Global
Lens type
Not Reported
HFOV
Not reported
VFOV
Not reported
Notes
Six landscape ring cameras; exposure is centred on the LiDAR sweep.

Camera 2

Model
Not reported
Setup
Mono
Count
1
Effective cameras
1
Modality
Rgb
Resolution
1550 x 2048
Rate
20 Hz
Shutter
Global
Lens type
Not Reported
HFOV
Not reported
VFOV
Not reported
Notes
Portrait front-centre ring camera; this is the seventh ring view.

Camera 3

Model
Not reported
Setup
Stereo
Count
1
Effective cameras
2
Modality
Rgb
Resolution
2048 x 1550
Rate
20 Hz
Shutter
Global
Lens type
Not Reported
HFOV
Not reported
VFOV
Not reported
Notes
One forward stereo system contains two physical cameras. All nine optical streams are documented at 20 fps.

IMUs

Not reported

GNSS

Not reported

LiDAR

Velodyne VLP-32C

Model or name
Velodyne VLP-32C
Count
2
Dimensions
3
Rate
10 Hz
Channels
32
Range
Not reported

Radar

Not reported

Additional sensors

Not reported

Automotive

Argoverse 2 Motion Forecasting

Two hundred fifty thousand eleven-second urban scenarios with planar actor histories, futures, headings and vector maps; no raw sensor streams.

Automotive

Argoverse 2 LiDAR

Twenty thousand thirty-second LiDAR scenarios with map-aligned ego poses and local vector maps, without imagery or object-box annotations.

Annotations

Reported annotation availability and task support.

Availability
Available
Format
Tracked 3D cuboids at 10 Hz; thirty native classes and twenty-six evaluation classes
Class count
30
Semantic segmentation Not reported
Instance segmentation Not reported
Object detection Available
Optical flow Not reported
Depth ground truth Not reported

Citation

Citation key and BibTeX kept at the end of the page for reference.

Citation key
wilson2021argoverse2

BibTeX

@INPROCEEDINGS { wilson2021argoverse2,
  author = {Benjamin Wilson and William Qi and Tanmay Agarwal and John Lambert and Jagjeet Singh and Siddhesh Khandelwal and Bowen Pan and Ratnesh Kumar and Andrew Hartnett and Jhony Kaesemodel Pontes and Deva Ramanan and Peter Carr and James Hays},
  title = {Argoverse 2: Next Generation Datasets for Self-driving Perception and Forecasting},
  booktitle = {Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks (NeurIPS Datasets and Benchmarks 2021)},
  year = {2021}
}

Dataset corrections

Notice a specific issue after reviewing the page?

Report an inaccuracy