USEFUL

mUltimodal Sensing and pErception For autonomUs vehicLes

A driving dataset with nine synchronised sensor streams (LiDAR, radar, RGB, thermal, SWIR and polarimetric) and 3D/2D box annotations, recorded in clear weather, rain, snow, fog, glare and at night.

Devkit · Download · Data format · License · Citation

At a glance

Release	v1.0
Sensor streams	9 (1 LiDAR, 2 radar, 6 cameras)
Recording logs / scenes	75 / 83
Annotated samples	10,953 (all keyframes, all sensors)
3D boxes	86,662, across 6,274 tracked instances
2D boxes	636,203, per camera
Object classes	12
Ego motion	GPS / INS / IMU / heading, per sample and as per-log CSVs
Size	≈ 230 GB (pick sensors and splits to download less)
Region	Catalonia (city, rural road, highway)
Captured	November 2024 – September 2025

Sensors

Channel	Modality	Format	Notes
LIDAR	LiDAR	.pcd	Beamagine L3CAM, MEMS quasi-solid-state, 60° × 20° FOV, XYZ + intensity + RGB
RADAR_LEFT	Radar	.pcd	Left corner, with Doppler velocity
RADAR_RIGHT	Radar	.pcd	Right corner, with Doppler velocity
NARROW	RGB	.png	Front, narrow FOV, 2064 × 1544
WIDE_LEFT	RGB	.png	Front-left, wide FOV, 2064 × 1544
WIDE_RIGHT	RGB	.png	Front-right, wide FOV, 2064 × 1544
LWIR	Thermal (long-wave IR)	.png	320 × 240
SWIR	Short-wave IR	.png	640 × 512
POLARIMETRIC	Polarimetric	.bmp	2448 × 2048 raw polarisation mosaic (0°/45°/90°/135°) → DoLP, AoLP, RGB

All channels are synchronised per sample and calibrated: calibrated_sensor gives every sensor's extrinsics (and, for cameras, intrinsics) relative to the ego vehicle.

Conditions and splits

Splits are made per scene, so no scene is shared between splits.

Split	Scenes	Samples
train	55	7,245
val	8	1,085
test	20	2,623

Weather	Scenes		Location	Scenes
clear	41		city	41
rain	10		road	28
snow	10		highway	14
night	10
fog	9
solar glare	3

Each log record also has a weather_conditions block: temperature, humidity, precipitation, irradiance, visibility, sun elevation, and day/twilight/night. These values come from the nearest weather stations and ERA5 reanalysis, not from sensors on the vehicle.

Annotations

3D boxes are given in the LiDAR / ego frame (centre, size, quaternion) with LiDAR point counts. 2D boxes are given per camera image, with a visibility level (0–40 %, 40–60 %, 60–80 %, 80–100 %). Instances are tracked across the frames of a scene.

Class	3D boxes
vehicle.car	37,095
sign	23,902
pedestrian	12,697
vehicle.truck	4,084
vehicle.motorcycle	3,650
bicycle	3,408
vehicle.bus	1,035
personal_mobility	408
animal	265
vehicle.construction	110
vehicle.motorcycle.crowd	8
bicycle.crowd	0

Privacy

Faces and licence plates are blurred in every image of the NARROW, WIDE_LEFT, WIDE_RIGHT, POLARIMETRIC and SWIR channels: 13,617 face and 53,716 plate regions in 32,400 images. The regions combine face and plate labels that were revised by hand with the detections of a YOLO11-m face/plate detector, which was run on every camera frame at a low confidence threshold (0.10) so that as few regions as possible are missed. Before release, every region was checked automatically against the original image, and random samples were checked by eye. LWIR (320 × 240 thermal) shows neither readable plates nor identifiable faces, so it is not blurred.

The images were re-anonymised on 2026-10-15. The first upload of this release (2026-10-14) blurred only the 53,185 regions of the earlier labels; please download the images again if you have that copy.

If you find a face or plate that is still recognisable, please open a discussion on this repository with the file name, and we will fix it. Do not try to re-identify anyone in the data.

Download

The data is split into one self-contained .tar per sensor × scene, under shards/<CHANNEL>/<split>/<scene>.tar. Each tar can be downloaded and extracted on its own, so you only fetch the sensors and splits you need.

pip install -U huggingface_hub

# Everything (~230 GB)
hf download GerardDMG/USEFUL --repo-type dataset --local-dir USEFUL

# Metadata + LiDAR + front camera only
hf download GerardDMG/USEFUL --repo-type dataset --local-dir USEFUL \
  --include "v1.0/*" --include "shards/LIDAR/*" --include "shards/NARROW/*"

# Only the validation split, all sensors
hf download GerardDMG/USEFUL --repo-type dataset --local-dir USEFUL \
  --include "v1.0/*" --include "shards/*/val/*" --include "shards/GPS/*" \
  --include "shards/HDG/*" --include "shards/IMU/*" --include "shards/INS/*"

Or from Python:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="GerardDMG/USEFUL",
    repo_type="dataset",
    local_dir="USEFUL",
    allow_patterns=["v1.0/*", "shards/LIDAR/*", "shards/NARROW/*"],
)

Downloads are checksummed by the Hub client and resume if interrupted.

Extract

Every tar unpacks into samples/<CHANNEL>/…. Extract them from the dataset root, in parallel and in any order:

cd USEFUL
find shards -name '*.tar' -print0 | xargs -0 -n1 -P8 tar -xf
rm -r shards   # optional, once extracted

Data format

After extraction:

USEFUL/                          <- dataroot
├── v1.0/                        <- metadata, 12 JSON tables
└── samples/
    ├── LIDAR/  RADAR_LEFT/  RADAR_RIGHT/                       .pcd
    ├── NARROW/  WIDE_LEFT/  WIDE_RIGHT/  LWIR/  SWIR/                 .png
    ├── POLARIMETRIC/                                           .bmp
    └── GPS/  HDG/  IMU/  INS/                                  one .csv per log

Sensor files are named {vehicle}-{datetime}__{CHANNEL}__{timestamp_us}.{ext}, for example u001-2024-11-19-12-46-25+0100__LIDAR__1732020526668479.pcd.

The metadata follows the nuScenes schema. Every record has a 32-character hex token, and tables point to each other by token:

Table	One record per…
log	recording session: vehicle, date, location, weather
scene	clip within a log: description, split, first/last sample
sample	synchronised multi-sensor keyframe (prev / next linked)
sample_data	sensor file belonging to a sample
ego_pose	GPS, INS, IMU and dual-antenna heading readings at a sample
sensor	sensor channel and modality
calibrated_sensor	sensor's extrinsics and intrinsics
sample_annotation	3D box of an instance in a sample
sample_annotation_2d	2D box of an instance in one camera image
instance	tracked object
category	object class
visibility	visibility level of a 2D box

Devkit

The USEFUL devkit loads the tables, resolves file paths, renders samples and scenes, exports mmdetection3d info files, and evaluates 3D detection.

git clone https://github.com/GDMG99/useful-devkit.git
pip install -e useful-devkit

from useful import USEFUL

usfl = USEFUL(version="v1.0", dataroot="USEFUL", verbose=True)
scene = usfl.scene[0]
sample = usfl.get("sample", scene["first_sample_token"])
print(usfl.get_sample_data_path(sample["data"]["LIDAR"]))

License

CC BY-NC 4.0: free for non-commercial use, with attribution.

Acknowledgements

The data format and devkit build on nuScenes and TruckScenes.

Citation

If you use USEFUL, please cite the dataset:

@misc{useful_dataset_2026,
  title        = {{USEFUL}: mUltimodal Sensing and pErception For autonomoUs vehicLes},
  author       = {DeMas-Giménez, Gerard and Subirana, Adrià and Bernal, Eduardo and García-Gómez, Pablo and Casas, Josep Ramon and Royo, Santiago},
  year         = {2026},
  version      = {v1.0},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/GerardDMG/USEFUL}},
  note         = {Dataset}
}