nvidia/tuv-data

TUV CCRs dataset

44 reconstructed CCRs (car-to-car rear stationary) clips from drive 5582a790-f497-43a4-9bfc-386f7e748965, with vehicle-under-test (VUT) target speeds of 30–50 km/h. The collection contains approximately 59.1 GiB of artifacts. Start with metadata/clips.csv: each row identifies one clip and provides its timestamps, labels, and file paths.

Every clip has a NuRec USDZ, RC log, five radar recordings with seek indexes, and a clip-specific IMU recording. 42 clips also have a matched GT CSV. Full-drive IMU and GPS recordings are stored once and referenced by all clips. GPS is available at drive level, not as a separate per-clip file.

Folder layout

README.md
metadata/
  clips.csv                 One row per clip: the primary catalog
  files.csv                 Artifact paths and byte sizes
examples/
  inspect_clip.py           Local, dependency-free clip inspection example
clips/<clip_id>/
  nurec/last.usdz            Neural reconstruction bundle
  recordings/
    roadcast_debug.log      RC log
    imu_vehicle.bin         Clip-specific IMU recording
    radar_*.bin             Five radar sensor recordings
    radar_*.bin.seek        Matching radar seek indexes
  ground_truth/V14_*.csv     Matched GT data; absent for two clips
drives/<drive_id>/
  recordings/
    imu_vehicle.bin         Full-drive IMU
    gps_vehicle.bin         Full-drive GPS

Radar positions are corner_front_left, corner_front_right, corner_rear_left, corner_rear_right, and front_center. Keep each .bin.seek beside its corresponding .bin. Original recording basenames are retained.

Start locally

Run from the dataset root:

python3 examples/inspect_clip.py --root .
python3 examples/inspect_clip.py --root . --clip-id 82aa0b4c-d23e-40ed-b5a5-86bdc5eecfcc

The example checks that the selected clip's paths exist and prints GT column names, units, and the first data row when GT is present. It does not decode the sensor binaries or download anything.

To load a GT CSV with pandas:

from pathlib import Path
import pandas as pd

root = Path('.')
catalog = pd.read_csv(root / 'metadata/clips.csv')
clip = catalog[catalog.gt_status == 'copied'].iloc[0]
gt_path = root / clip.gt_path
units = pd.read_csv(gt_path, nrows=1).iloc[0].to_dict()
gt = pd.read_csv(gt_path, skiprows=[1], low_memory=False)
print(gt[['System Time', 'Speed']].head())

GT CSV row 1 contains column names, row 2 contains units, and subsequent rows contain measurements. Files are UTF-8 with BOM. They were converted from the original tab-delimited text exports; the vendor preamble and footer were excluded. Data values were retained. System Time is in seconds and Speed is in km/h. GT captures retain their full original duration; they have not been trimmed or resampled to the reconstructed clip window.

Load from Hugging Face after upload

Replace ORG/DATASET with the actual dataset repository ID. Install datasets and huggingface_hub as needed. The README configuration exposes only the clip catalog to the Dataset Viewer; GT CSVs have their own measurement schema.

from datasets import load_dataset
from huggingface_hub import snapshot_download

catalog = load_dataset('ORG/DATASET', name='clips', split='train')
clip = next(row for row in catalog if row['gt_status'] == 'copied')
local_root = snapshot_download(
    repo_id='ORG/DATASET',
    repo_type='dataset',
    allow_patterns=['metadata/clips.csv', f"clips/{clip['clip_id']}/*"],
)
print(local_root, clip['gt_path'])

This downloads one clip's artifacts, including its NuRec. To download only GT, use allow_patterns=[clip['gt_path']]. To obtain shared GPS or full-drive IMU, add clip['drive_gps_path'] or clip['drive_imu_path'] to the patterns. The train split is a catalog convention; no training/validation/test partition or independence between clips is implied. All clips share one drive.

Repository configuration follows the Hugging Face dataset structure documentation. This folder is prepared locally; no repository has been created or uploaded. No redistribution license has been specified for this collection.

Catalog fields and clocks

All *_path fields are relative to the dataset root. An empty gt_path means unavailable GT, not an empty measurement file.

Fields	Meaning
clip_id, drive_id, event_id, event_number	Clip, source drive, and source event identifiers
scenario, modifier	CCRs scenario and original event modifier tag
vut_kmh	Target ego speed from the event, not measured clip-start speed
event_overlap_pct, gt_overlap_pct	Event description overlap and matched GT workbook overlap
event_run, gt_run	Original event run label and matched workbook run
overlap_label_discrepancy	Whether the two overlap labels differ
start_time_ptp_us, end_time_ptp_us	Registered clip bounds in the source drive's PTP microsecond clock
duration_s	(end_time_ptp_us - start_time_ptp_us) / 1e6
gt_status	copied or missing_source
nurec_version	Reconstruction version from the download location
*_path	Direct paths to clip artifacts or shared drive recordings

Preserve timestamp columns as 64-bit integers. Do not interpret these values as Unix timestamps. The catalog contains the registered clip bounds used for the downloads; no additional start/end buffer has been applied during packaging.

GT System Time uses a different clock. Do not directly compare it with the PTP timestamps. The file matching used timing and speed-trace checks; this is not certified sample-level synchronization. No clock correction or indicated-speed bias correction has been applied to the GT data.

Known gaps and label discrepancies

GT is unavailable for:

• 05f46bdf-531f-469e-8676-780feea186f1 — source event S34; 30 km/h, 125% overlap, run 3.
• ccf48470-57e4-4778-84c3-58e73984cb6b — source event S42; 30 km/h, 50% overlap, run 2.

Three clips have event overlap labels of 25% while the time-matched GT workbook labels are 0%, at 40 km/h:

Clip	Matched GT
331de149-2105-4c0c-abe9-87d70a7ffcba	V14_T123_R1.csv
e217710d-b043-42a6-b35c-9330b705f170	V14_T123_R2.csv
cd655365-3026-46b9-acbc-eb239ce2148d	V14_T123_R3.csv

Both labels are retained so consumers can choose explicitly. The matching audit verified all 42 CSV conversions against their source data and checked file association with timing and speed traces. This is not independent verification of the physical overlap or the scenario outcome.

Formats

• NuRec: last.usdz is the original downloaded reconstruction bundle. Use compatible NuRec/USD tooling; it may contain model/configuration assets beyond a displayable USD scene.
• RC, radar, IMU, GPS: native recording formats. Compatible recording decoders are required and are not included here. These .bin files are not documented as raw NumPy arrays.
• GT: directly readable CSV measurements, with the units row described above.
• File inventory: metadata/files.csv lists artifact paths and byte sizes. Packaging moved files without re-encoding their contents. Finder metadata was excluded.