List of Datasets

Automotive

nuScenes

Multimodal 3D detection and tracking for autonomous driving

1000 selected 20-second driving scenes in Boston and Singapore with six cameras, five radars, one lidar, 3D boxes, attributes, semantic maps and estimated ego localization.

Metadata: Complete Metadata curator reviewed Released 2019

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

Quick facts

Vehicle type
Automotive
Environment
Outdoor
Data origin
Real-world
Sensor count
13
Ground truth
Available
Calibration
Parameters Only
Annotations
Available

Sensor overview

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

Cameras

6 reported

IMUs

1 reported

GNSS

Not reported

LiDAR

1 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 Not reported Not reported Not reported Full 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
Accuracy and notes
Accuracy notes
Main text claims localization errors ≤10 cm; supplement says no ground-truth localization data and offers only qualitative reconstruction. Not an independent accuracy guarantee.
Reference notes
Ego-pose metadata contains XYZ translation and quaternion rotation in the global frame. Poses are estimated from lidar-map localization; CAN pose outputs are another distribution of vehicle state, not a second independent reference. Base pose cadence is not a single fixed release-wide rate.
3D position + heading Manual annotation Not reported Not reported 2 Hz Partial · Objects With at Least One Lidar or Radar Point at 2 Hz Keyframes. 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
Accuracy and notes
Reference notes
Human 3D cuboids with position, dimensions and yaw, semantic class/attributes across each scene. Ten detection and seven tracking classes are task mappings from native 23. Reference follows sensor-visible labeled objects, not externally tracked targets.

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

Calibration parameters in relational metadata. Extrinsics refer to rear axle midpoint; camera cube target and intrinsic board, lidar laser liner and road radar yaw fitting. Updated around twice weekly over six months. Raw calibration measurements are unestablished.

Known or intentional limitations

Reported constraints and characteristics to check before using the dataset.

Selected 20-second scenes are not continuous release of all 84 source logs or 242 km.

Severe class imbalance, night visibility challenges and rain on windshields affect annotations.

Ego localization lacks independent accuracy validation; paper error claims are qualified by the supplement.

Images have blurred faces/license plates. Later lidarseg and panoptic labels extend the paper-era box annotations and use task-specific taxonomies.

Download requires an account and acceptance of the dataset terms. CAN data are an experimental extension.

Cameras

Camera 1

Model
Not reported
Setup
Mono
Count
6
Effective cameras
6
Modality
Rgb
Resolution
1600 x 900
Rate
12 Hz
Shutter
Not Reported
Lens type
Not Reported
HFOV
Not reported
VFOV
Not reported
Notes
Auto-exposed JPEG images. Front/side FoV reported as 70 degrees and rear as 110 degrees, without an axis convention. Camera models/shutter type unreported.

IMUs

CAN bus IMU

Model
Not reported
Count
1
Accel / gyro
100 / 100 Hz

GNSS

Not reported

LiDAR

LiDAR 1

Model or name
LiDAR 1
Count
1
Dimensions
3
Rate
20 Hz
Channels
32
Range
70 m

Additional sensors

Radar

Count
5
Rate
13 Hz
Notes
77 GHz FMCW; reported device range ≤250 m and velocity accuracy ±0.1 km/h. Doppler velocity measurements are distinct from labeled object velocities.

Other sensor

Name
CAN bus chassis measurements
Count
1
Notes
Remaining CAN chassis measurements include steering, wheel speeds and other vehicle state. Rates vary by message; physical hardware counts are unestablished. IMU outputs are represented separately.

Annotations

Reported annotation availability and task support.

Availability
Available
Format
[relational_database Point-wise lidarseg semantic labels and panoptic semantic/instance labels for annotated keyframes.]
Class count
Not reported
Semantic segmentation Available
Instance segmentation Available
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
caesar2020nuscenesv5

BibTeX

@misc{caesar2020nuscenesv5,
  title={nuScenes: A multimodal dataset for autonomous driving},
  author={Caesar, Holger and Bankiti, Varun and Lang, Alex H. and Vora, Sourabh and Liong, Venice Erin and Xu, Qiang and Krishnan, Anush and Pan, Yu and Baldan, Giancarlo and Beijbom, Oscar},
  year={2020},
  eprint={1903.11027},
  archivePrefix={arXiv},
  url={https://arxiv.org/abs/1903.11027v5},
  note={Analyzed version 5, dated 5 May 2020}
}

Dataset corrections

Notice a specific issue after reviewing the page?

Report an inaccuracy