Hugging Face's logo

Datasets:
natix-network-org
/
natix-multi-camera-driving-dataset

Modalities:
Video
Geospatial
Languages:
English
Size:
1K<n<10K
Tags:
autonomous-driving
computer-vision
video
dashcam
multi-camera
geospatial
License:
Dataset card Data Studio Files
xet
 Community

You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content. Once your access request is approved, you will be provided with the credentials required to download the full dataset.

Log in or Sign Up to review the conditions and access this dataset content.

NATIX Multi-Camera Driving Dataset

Summary

The NATIX Multi-Camera Driving Dataset is a real-world driving dataset. Unlike fleet-collected datasets that rely on trained safety drivers operating standardized vehicles, NATIX sources its data from a decentralized vehicle network consisting of everyday, non-expert drivers, resulting in more natural driving behavior and a broader distribution across real-world scenarios.

NATIX multi-camera driving sample

NATIX plans to release 2,000+ hours of data as part of this initiative. This first public release contains 100 hours of multi-camera driving footage with GPS/GNSS, camera calibration, and trip-level metadata collected across multiple countries, road types, weather conditions, and driving environments. The dataset is designed to support research in Physical AI, autonomous driving, world models, robotics, mapping, and end-to-end driving systems.

Note that the GPS data provided in this dataset is collected via a consumer-grade sensor installed on our data collection device. NATIX is also collecting telemetry data, including but not limited to automotive-grade GPS, pedal position, wheel angle, and more. We will publish a telemetry-included dataset soon.

This dataset supports research on:

  • End-to-end driving models
  • Vision-Language-Action (VLA) models
  • World foundation models
  • Multi-camera perception
  • Mapping
  • Scenario mining via VLM
  • Driving scene understanding and simulation in graphics-based simulators

We strongly support open research and initiatives. If you are a researcher, academic institution, or open-source project interested in accessing larger portions of the dataset than publicly available, we'd love to hear from you, and we're happy to support and help accelerate research in Physical AI and autonomous driving. For more information, please contact us at dataset@natix.io.

If you are using NATIX data for any applications, please contact us to feature your project here.

Each trip contains:

  • Footage, recorded by Tesla dashcams and segmented into approximately 1-minute .mp4 clips
  • GPS metadata, delivered per .mp4 clip as CSV
  • Trip-level metadata, including Trip Insight and Fixed Metadata

Important Notes

  • This dataset is anonymized: faces and license plates blurred; known sensitive/military areas removed. However, real-world crowd-sourced data can contain unexpected cases. In case anything sensitive, personal, restricted, incomplete, or inconsistent appears in the dataset, the consumer of this dataset is responsible for immediately notifying NATIX at dataset@natix.io and deleting the affected data from their local copy.
  • Metadata fields are best-effort estimates, especially values derived from third-party APIs or baseline calculations, such as weather, temperature, estimated distance, and estimated average speed. Treat these fields as guidance, not ground truth.

Dataset Statistics (Full dataset)

Metric Value
Dataset name NATIX Multi-Camera Driving Dataset
Release identifier natix-multi-camera-driving-dataset
Trips 2,736
Countries 2: Switzerland, United States
Total duration 100.0 hours / 6,001 minutes
MP4 clips 30,902
Total files 114,614
Total size 1,283.58 GB
4-camera footage 2,551 minutes
6-camera footage 3,450 minutes
Switzerland 30 minutes
US / California 4,209 minutes
US / Colorado 18 minutes
US / Florida 1,278 minutes
US / Georgia 66 minutes

Access and Download the Full Dataset

The Hugging Face repository is only a 20-minute sample (dataset-sample/ folder), which contains 6 complete trips selected from Switzerland and the United States. The full dataset is hosted externally on Cloudflare R2 (larger than 1 TB)

This dataset is gated. To download the dataset, please request access on this page. Once your access request is approved, you will be provided with the ACCESS_KEY and SECRET_KEY credentials required for the download script below.

Note: you need 1,283.58 GB of free storage

Install dependencies:

pip install boto3

Download the full dataset:

import boto3
from pathlib import Path

ENDPOINT_URL = "https://e613d208b194c3a6749f5cc0a1ca5510.eu.r2.cloudflarestorage.com"
BUCKET_NAME  = "natix-prod-100h-open-source"
ACCESS_KEY   = "PROVIDED_ACCESS_KEY"    # provided after access request is approved
SECRET_KEY   = "PROVIDED_SECRET_KEY"    # provided after access request is approved
DEST_DIR     = Path("./dataset")

s3 = boto3.client(
    "s3",
    endpoint_url=ENDPOINT_URL,
    aws_access_key_id=ACCESS_KEY,
    aws_secret_access_key=SECRET_KEY,
)

paginator = s3.get_paginator("list_objects_v2")
for page in paginator.paginate(Bucket=BUCKET_NAME):
    for obj in page.get("Contents", []):
        key       = obj["Key"]
        dest_path = DEST_DIR / key
        # Creating parent directories as needed
        dest_path.parent.mkdir(parents=True, exist_ok=True)
        print(f"Downloading: {key}")
        s3.download_file(BUCKET_NAME, key, str(dest_path))

Dataset Overview

  • Footage: recorded via vehicle surround-view camera system and segmented into approximately 1-minute .mp4 files.

    • 4-camera trips: FRONT, REAR, LEFT, RIGHT.
    • 6-camera trips: FRONT, REAR, LEFT_REPEATER, RIGHT_REPEATER, plus two pillar cameras: LEFT_PILLAR, RIGHT_PILLAR.
    • Camera FOV (Field of View): Front ~50°; Rear ~140°; Side ~90°; Side Pillar ~90°.
    • Resolution depends on the vehicle type and camera. Observed examples include
      • 1280 x 960 across all cameras in some 4-camera trips
      • 1448 x 938 for the standard cameras in some 6-camera trips, and,
      • in some 6-camera trips, a natively larger front camera, for example, 2896 x 1876.
    • All faces and license plates are blurred. In addition, 14% of the bottom of all rear-camera footage is blurred/anonymized.

  • Time-series metadata

    • GPS metadata, 1-10 Hz: Per-camera, frame-matched location and motion metadata.

  • Trip-level metadata

    • Trip Insight: high-level aggregated metadata about the trip, including duration, estimated distance, estimated average speed, per-minute weather, road type, location context, and camera availability.
    • Fixed Metadata: values that do not change during a trip, including vehicle/platform information, camera intrinsics, camera extrinsics, and sensor positions.

Artifacts included in this Dataset

Component Location Availability
Footage (.mp4, per camera) Camera folder Included
GPS metadata (.csv, per camera) Camera folder Included
trip_insight.json Trip root Included
fixed_metadata.json / .csv Trip root Included
Schema files (telemetry_data.proto, trip_metadata.proto) Trip root Included
mapping.txt Segment folder Expected. Contact NATIX if missing or inconsistent.
trip_manifest.jsonl / Included. Trip-level manifest for all trips
dataset-sample/ / Included. 6 complete trips from Switzerland and the United States

Dataset Folder Structure 

📁 dataset/
├── 📁 Country/[State]/
│   └── 📁 <trip-id>/
│       ├── 📁 HH-MM-SS/
│       │   ├── 📁 FRONT_FOLDER/
│       │   │   ├── FRONT_<YYYY-MM-DD>_<HH-MM-SS>.mp4
│       │   │   ├── FRONT_<YYYY-MM-DD>_<HH-MM-SS>.csv
│       │   │   └── FRONT_<YYYY-MM-DD>_<HH-MM-SS>.mcap
│       │   ├── 📁 REAR_FOLDER/
│       │   │
│       │   │   # 4-camera trips additionally contain:
│       │   ├── 📁 LEFT_FOLDER/
│       │   ├── 📁 RIGHT_FOLDER/
│       │   │
│       │   │   # 6-camera trips additionally contain:
│       │   ├── 📁 LEFT_REPEATER_FOLDER/
│       │   ├── 📁 RIGHT_REPEATER_FOLDER/
│       │   ├── 📁 LEFT_PILLAR_FOLDER/
│       │   ├── 📁 RIGHT_PILLAR_FOLDER/
│       │   │
│       │   └── mapping.txt
│       │
│       ├── 📁 HH-MM-SS/
│       │   └── ...
│       │
│       ├── trip_insight.json
│       ├── fixed_metadata.json
│       ├── fixed_metadata.csv
│       ├── fixed_metadata.mcap
│       ├── telemetry_data.proto
│       └── trip_metadata.proto

Data Definitions

Trip Manifest

The manifest contains one JSON object per trip folder, not one JSON object per individual file. Each row describes a self-contained trip folder and includes the trip prefix, country or region, duration, camera configuration, file count, total size, counts for per-camera data files, and segment timestamps. The complete trip-level manifest is provided as: trip_manifest.jsonl

Field Description
trip_id Trip folder identifier, including the continuous-piece suffix where applicable
country Country of the trip
state State or region when available, otherwise null
prefix Relative path to the trip folder under the R2 dataset root
duration_min Trip duration in minutes
cam_count Number of cameras in the trip configuration
cameras Camera folders present for the trip
file_count Total number of files inside the trip folder
size_bytes Total trip folder size in bytes
formats Count of per-camera data files by extension, such as mp4 and csv
timestamps Segment start-time folders included in the trip

GPS Metadata

Each GPS Metadata file is located in its camera folder and is named like the corresponding .mp4 file, for example, FRONT_2025-11-12_21-07-33.csv. It contains frame-matched GPS position and motion metadata.

Field Description
timestamp ISO-8601 timestamp with millisecond precision
frame_number Frame index synced to the .mp4, starting at 1
GPS_latitude_deg, GPS_longitude_deg Latitude and longitude
horizontal_accuracy_m Location uncertainty in meters
speed_mps Estimated speed in meters per second, from the location data
velocity_north_mps, velocity_east_mps Velocity components toward north and east in meters per second
heading_deg Heading in degrees, clockwise, 0 = north, calculated from position updates
heading_accuracy_deg Heading uncertainty in degrees
image_direction Direction the camera faces in this frame, in degrees, clockwise, 0 = north; calculated from heading and camera mounting direction

Notes 1- Missing values are encoded as the literal string na. 2- Rows may skip frame numbers on any camera, including the front camera. When no GPS update is available for a footage frame, that frame may have no row. Do not assume one row per frame.

Fixed Trip Metadata

At the root of each trip folder, static information that does not change during the trip is provided in multiple equivalent formats: fixed_metadata.json and fixed_metadata.csv.

Top-level fields

Field Description
version Schema version
trip_identifier Trip UUID, without the _<n> piece suffix
frame_width, frame_height, frame_mp Footage frame size in pixels and megapixels; see the note below
vehicle_make, vehicle_model Vehicle make and model, for example, Tesla / model3 or modely
platform_type_video Source platform of the footage, for example, Tesla
reference_frame Reference frame of all extrinsics, for example, ground_nominal

Front camera resolution note frame_width and frame_height describe the trip standard cameras. In some 6-camera trips, the front camera is natively larger, for example, 2896 x 1876 instead of 1448 x 938. Its intrinsic parameters, such as fx, fy, cx, and cy, are already expressed at the native front-camera size. Use them as shipped together with the actual pixel size of the front footage. NATIX does not resize the footage.

device_extrinsics - one entry per device. Camera entries contain:

Field Description
device_name e.g. camera_front, camera_rear; camera_left, camera_right in 4-camera trips; camera_left_repeater, camera_right_repeater, camera_left_pillar, camera_right_pillar in 6-camera trips
fx, fy, cx, cy Camera intrinsics, estimated, in pixels
k1, k2, k3, p1, p2 Radial and tangential distortion, estimated
r11 ... r33 3 x 3 row-major rotation matrix: orientation of the camera relative to ground_nominal
tx, ty, tz Translation of the device relative to ground_nominal, based on usual device locations
field_of_view_hor_deg, field_of_view_ver_deg Horizontal and vertical field of view in degrees

Trip Insight

At the root of each trip folder, trip_insight.json gives an overview of the whole trip. For split trips, it describes the standalone trip piece.

Trip-level fields

Field Description
startEpochMs, endEpochMs Trip start and end time in Unix epoch milliseconds
duration Trip duration in milliseconds
estimatedDistance Estimated distance traveled in kilometers
estimatedAverageSpeed Estimated average speed over the trip in km/h
firstLocation Object with latitude and longitude of the trip first GPS position
timezone IANA timezone string, for example, America/New_York
minutes Array of per-minute objects, each keyed by YYYY-MM-DD_HH-MM

Per-minute fields

Field Type Description
weather list Observed weather conditions, for example, ["Clear"]
temperature list Temperature in C, for example, [7.4]
timeOfDay list Time-of-day category, for example, ["day"], ["dusk"]
roadType list OSM-based road type, for example, ["motorway"], ["residential"]
country string Country name
region string Region or state
place string City or town
district string Administrative district
postcode string Postal code
locality string Locality sub-area
neighborhood string Neighborhood name, may be empty
address string Street address or area name
cameraCount number Number of cameras present in this minute
footageCount number Number of cameras with available footage in this minute
camera flags boolean Whether footage exists per camera

The camera-existence flags match the trip camera configuration:

  • 4-camera trips: frontCameraExists, rearCameraExists, leftCameraExists, rightCameraExists
  • 6-camera trips: frontCameraExists, rearCameraExists, leftRepeaterCameraExists, rightRepeaterCameraExists, leftPillarCameraExists, rightPillarCameraExists

Special Considerations

General Footage and File Notes

  1. Footage may have different frame rates and durations, even within the same segment.
    • Different durations, same frame rate: Front - 00:01:00.21 @ 36.02 fps; Rear - 00:01:00.00 @ 36.02 fps.
    • Different durations, different frame rates: Front - 00:01:00.06 @ 36.03 fps; Rear - 00:01:00.36 @ 34.62 fps.
  2. Split trips use the <trip-id>_<n> folder format. Each piece contains continuous minutes and has its own metadata.
  3. If footage, metadata, or camera-folder contents appear missing or inconsistent, please contact NATIX so the dataset can be reviewed.

GPS Data Processing

  1. GPS rows can skip frame numbers on any camera, including the front camera. Do not assume one metadata row per frame.
  2. GPS data is first aligned with the front footage. Because camera durations and frame rates can differ, other cameras may have different row counts or different matched metadata per frame.
  3. Time sync error between vehicle camera/video data and GPS data is expected to be 0-1 seconds, and in extreme cases up to 3 seconds.
  4. When GPS updates arrive too quickly, such as in two consecutive frames, heading_deg and speed_mps may be 0 because they are calculated from position updates.

Trip Insight Estimate Notes

Some Trip Insight values, such as estimatedDistance, estimatedAverageSpeed, weather, and temperature, are best-effort estimates derived from third-party APIs or baseline calculations. Treat them as guidance, not ground truth.

License and Usage Terms

This dataset is released by NATIX under the NATIX Data RAIL-NC License, a responsible-AI data license adapted from the BigScience Open RAIL-M License.

Under this License, the dataset may be used for non-commercial purposes only, subject to the use-based restrictions set out in the License, and may not be redistributed or made available to third parties. The License is granted for a limited term. Commercial use is NOT permitted without separate written permission from NATIX. For the full license text, see LICENSE.md.

Disclaimer This dataset is provided "as is" and "as available", without warranties of any kind, whether express or implied, including but not limited to warranties of merchantability, fitness for a particular purpose, accuracy, or non-infringement. While faces and license plates have been blurred and known sensitive areas removed, NATIX does not warrant that anonymization is complete or that all sensitive areas have been excluded. NATIX makes no guarantees regarding the completeness, reliability, or correctness of the data and provides no support, maintenance, or updates. Use of this dataset is entirely at the consumer's own risk, and NATIX shall not be liable for any damages or losses arising from its use.

Attribution and Citation

There is no separate paper required for citation. If you use this dataset, please credit NATIX and link to the dataset page.

Recommended attribution:

 NATIX Multi-Camera Driving Dataset. 2026. Available on Hugging Face Hub. NATIX Website: https://www.natix.network/

BibTeX:

@misc{natix2026_multi_camera_driving_dataset,
  title        = {NATIX Multi-Camera Driving Dataset},
  author       = {{NATIX}},
  year         = {2026},
  publisher    = {NATIX},
  howpublished = {Hugging Face Hub},
  url          = {https://huggingface.co/datasets/natix-network-org/natix-multi-camera-driving-dataset},
  note         = {Multi-camera driving dataset with telemetry metadata. Website: https://www.natix.network/}
}

Contact Us

We check every dataset before release, but real-world crowd-sourced data can contain surprises. If anything in the dataset looks sensitive, incomplete, inconsistent, or unexpected, please contact NATIX so it can be reviewed.

NATIX has built a unique, large-scale multi-camera driving dataset crowd-sourced from vehicles' cameras globally. This data is currently being used by various physical AI players supporting world foundational models, end-to-end (E2E) driving models, and simulation-based workflows for training, testing, and validation. For more info, you can contact NATIX directly at dataset@natix.io.
Downloads last month
4
Total file size:
4.81 GB