shchon11/TCar_hyu

T-Car nuScenes, full sensor set — release 1015

Driving data from the T-Car test vehicle with every sensor it records — seven cameras, five LiDARs, the front radar, RTK INS / GNSS, vehicle CAN — in the nuScenes v1.0-trainval format, with the camera and roof-LiDAR calibration.

554 scenes of 20 s (3.08 h) from 41 recordings · 22,160 key-frame samples · 3,318,262 sensor frames · 1435 GB to download in 268 sensor tars + 1 table tar (2001 GB unpacked). No 3D box annotations (sample_annotation is empty).

Download and use

1. Request access on this page (the dataset is gated), and log in: hf auth login.
2. Download this release and put it together (assemble.py checks every tar's sha256 and extracts it):

pip install -U huggingface_hub
hf download shchon11/TCar_hyu --repo-type dataset --revision 1015 --local-dir tcar_tars
python tcar_tars/assemble.py --tars tcar_tars --out tcar_nuscenes

from nuscenes.nuscenes import NuScenes
from nuscenes.can_bus.can_bus_api import NuScenesCanBus
nusc = NuScenes(version="v1.0-trainval", dataroot="tcar_nuscenes")
can = NuScenesCanBus(dataroot="tcar_nuscenes")      # pose, ms_imu, meta per scene

Some recordings only: download the tables and the recordings you want, and let assemble.py skip the rest (the tables still list every scene; the files of the others are just absent):

hf download shchon11/TCar_hyu --repo-type dataset --revision 1015 --local-dir tcar_tars \
    --include "manifest.json" --include "assemble.py" --include "meta/*" \
    --include "sensors/A-1_*/*" --include "sensors/B-2_*/*"
python tcar_tars/assemble.py --tars tcar_tars --out tcar_nuscenes --allow-missing

Updating: Coming from 1002: run the two commands above in the same directories — only the tars new in 1015 are downloaded and extracted. Nothing already extracted is changed; the tables are replaced. Add --delete-tars to assemble.py to remove each tar after extracting it (then hf download fetches it again next time).

Releases and versioning

Each release is a branch of this repository (--revision <name>), a complete dataset on its own; a release is never changed after it is published. main always holds the latest release (a copy, stored once), so hf download without --revision gets it; give --revision to stay on one release. Later releases only add data:

• sensors/<recording>/<recording>.partNN.tar.gz — the sensor files (samples/, sweeps/ with the .time.bin next to each LiDAR frame, and ext/<scene>/) of a group of whole scenes of one recording, at most 10 GB unpacked (gzip: the LiDAR files shrink by about a third, the JPEGs not). Once published, a tar never changes and is carried into every later release (stored once).
• meta/TCar_meta.tar.gz — the tables (v1.0-trainval/), can_bus/, maps/, calibration/, the per-recording *.import.json (how each recording was converted) and curation/. The only file that changes between releases.
• manifest.json — every tar with its recording, scene tokens, files, bytes and sha256, and the release history.
• Scene tokens never change. Scene names (scene-0001, …) are the nuScenes train split names in recording order; the names of released scenes do not change either. All scenes are in the train split.

release	date	recordings added	scenes added	scenes in total	parser commit
1002	2026-10-09	A-1, A-2, A-3, A-4, A-5, A-6, A-7, A-8, A-9, A-10, B-1, B-2, B-3, B-4, B-5, B-6, B-7, D-1, D-2, D-4, D-5	268	268	3ec1e1e
1015	2026-10-16	C-1, C-2, C-3, C-4, C-5, C-6, C-7, C-8, D-3, E-1, E-2, E-3, E-4, E-5, E-6, F-1, F-2, F-3, F-4, F-5	286	554	3b68a5f (+local changes)

Contents

course	dates	recordings	scenes	time	GB
A	2026-09-28	10	119	40 min	327
B	2026-09-29	7	103	34 min	267
C	2026-10-06	8	133	44 min	364
D	2026-09-29	5	64	21 min	150
E	2026-10-06	6	58	19 min	142
F	2026-10-06	5	77	26 min	184

channel	sensor	data	rate	frames
CAM_BACK	camera	1920×1080 JPEG, as recorded	29 Hz	323,465
CAM_BACK_LEFT	camera	1920×1080 JPEG, as recorded	29 Hz	323,465
CAM_BACK_RIGHT	camera	1920×1080 JPEG, as recorded	29 Hz	323,465
CAM_FRONT	camera	1920×1080 JPEG, as recorded	29 Hz	323,465
CAM_FRONT_LEFT	camera	1920×1080 JPEG, as recorded	29 Hz	323,465
CAM_FRONT_RIGHT	camera	1920×1080 JPEG, as recorded	29 Hz	323,465
CAM_TRAFFIC	camera	1920×1080 JPEG, as recorded	29 Hz	323,465
LIDAR_BOTTOM_FRONT	lidar	point cloud (.pcd.bin + .time.bin)	10 Hz	108,584
LIDAR_BOTTOM_LEFT	lidar	point cloud (.pcd.bin + .time.bin)	10 Hz	108,584
LIDAR_BOTTOM_REAR	lidar	point cloud (.pcd.bin + .time.bin)	10 Hz	108,584
LIDAR_BOTTOM_RIGHT	lidar	point cloud (.pcd.bin + .time.bin)	10 Hz	108,584
LIDAR_TOP	lidar	point cloud (.pcd.bin + .time.bin)	10 Hz	108,584
RADAR_FRONT	radar	detections (.pcd, nuScenes radar fields)	20 Hz	215,935

Key frames (samples/) are 2 Hz, chosen where LIDAR_TOP and the six standard cameras are synchronised (the same frames as the standard set). CAM_TRAFFIC, the bottom LiDARs and the radar join a key frame when their nearest frame is close enough (25 ms for the camera, half a frame period otherwise); every other frame is a sweep (sweeps/). The images are the JPEGs as recorded: not rectified.

• LiDAR — .pcd.bin: five float32 per point (x, y, z, intensity, ring) in the sensor frame, as in nuScenes. Next to each one, <same name>.time.bin holds each point's time: float32 seconds relative to the frame's sample_data timestamp, one per point in the same order. LIDAR_TOP is the roof RoboSense Ruby 128; LIDAR_BOTTOM_FRONT / _REAR are the M1s, LIDAR_BOTTOM_LEFT / _RIGHT the Bpearls.
• Radar — RADAR_FRONT is the Continental ARS548's detection point cloud (/radar/PointCloudDetection) as nuScenes radar .pcd, in the sensor frame. The full detection list, object list and status are in ext/.
• ext/<scene>/<topic>.json (topic / → __) — every other topic of the recording, for the scene ± 0.5 s, as a JSON array of {"utime", "bag_utime", "msg"}: utime is the message header stamp (the bag time when there is none) in the same microseconds as sample_data, msg the message fields as recorded (their own units; constants omitted). _index.json lists each file's topic and ROS message type. About 240 MB per scene, 51 topics:
  • INS / GNSS (/novatel/oem7/*): inspva, inspvax, corrimu, insstdev, bestpos, bestgnsspos, bestutm, bestvel, ppppos, heading2, rxstatus, time, odom, and oem7raw (message_data is the receiver's binary log; concatenated in order it is a file NovAtel tools read, with logs the decoded topics lack, e.g. RAWIMUSX);
  • vehicle: decoded CAN /bsw/vehicle_can, /bsw/vehicle_can_extended, raw CAN frames /can_rx0, /can_rx1 (CAN FD), /can_tx0, /app/loc/vehicle_state, /bsw/ad_state, the mode switches (/bsw/*_mode_sw), /hmi/logging_status;
  • radar: /radar/DetectionList (all 800 slots as sent; only the first list_numofdetections are this cycle's detections), /radar/ObjectList, /radar/PointCloudObject, /radar/DirectionVelocity, /radar/Status, and the corner / vehicle radar objects (/bsw/corner_radar_objects*, /bsw/vehicle_radar_objects);
  • the cameras' camera_info (the driver's placeholder, not the calibration), the LiDARs' rslidar_status, and diagnostics /diag/ptp, /diag/disk.
• can_bus/ — the nuScenes CAN bus expansion (pose, ms_imu, meta) from the INS, as in the standard set.

Calibration

calibration/ is the calibration of 2026-09-23 (tcar_calib_20260923; its README.md documents the model and the accuracy): the seven cameras and LIDAR_TOP. Its values are also in calibrated_sensor, the same for every recording:

• translation, rotation — sensor → ego (T_ego_cam, T_ego_lidar). The ego frame is the NovAtel output point (x forward, y left, z up), the frame ego_pose places.
• camera_intrinsic — fx, fy, cx, cy of the Kannala-Brandt (OpenCV cv2.fisheye) model of the raw image. The distortion is in extra keys of the record: camera_model ("kannala_brandt"), camera_distortion (k1…k4), image_size, time_offset_s and rolling_shutter_readout_s (exposure time of each image row), and windshield — the file of the windshield refraction field, a smooth pixel offset added after the lens model (up to 3.6 px at 10 m without it).
• calibration — the calibration's name; null for the bottom LiDARs and the radar, which are not calibrated (identity placeholder).

The devkit's own projection (view_points) is a plain pinhole and ignores the distortion; to project LiDAR into the images use calibration/project.py (load_camera, project_lidar), which applies the lens model and the windshield field.

Coordinate frames

ego_pose is in the location's global frame, as in nuScenes: metres, x east, y north, z up, on the tangent plane at 37.20° N, 126.83° E (height 0 on the WGS84 ellipsoid). log.json names it in the extra key global_frame (enu@37.200000,126.830000,0.000). Ego poses and the CAN bus pose come from the NovAtel INSPVA solution (100 Hz, RTK), placed at the receiver's GPS measurement time.

Recordings

recording	date	scenes	tars	GB	first release
A-1_2026-09-28-14-40-58	2026-09-28	10	5	26.9	1002
A-2_2026-09-28-14-45-11	2026-09-28	10	5	27.7	1002
A-3_2026-09-28-16-32-23	2026-09-28	12	6	32.6	1002
A-4_2026-09-28-14-51-38	2026-09-28	10	5	27.9	1002
A-5_2026-09-28-14-55-45	2026-09-28	12	6	33.3	1002
A-6_2026-09-28-15-02-17	2026-09-28	15	8	41.9	1002
A-7_2026-09-28-15-08-17	2026-09-28	12	6	31.6	1002
A-8_2026-09-28-15-13-01	2026-09-28	10	5	27.0	1002
A-9_2026-09-28-15-19-08	2026-09-28	12	6	34.0	1002
A-10_2026-09-28-16-03-14	2026-09-28	16	8	44.3	1002
B-1_2026-09-29-11-30-09	2026-09-29	12	6	31.8	1002
B-2_2026-09-29-11-37-49	2026-09-29	10	5	27.1	1002
B-3_2026-09-29-11-42-12	2026-09-29	17	9	45.6	1002
B-4_2026-09-29-11-50-32	2026-09-29	12	6	30.9	1002
B-5_2026-09-29-11-55-19	2026-09-29	13	7	33.0	1002
B-6_2026-09-29-12-02-16	2026-09-29	24	11	61.1	1002
B-7_2026-09-29-12-19-37	2026-09-29	15	8	38.1	1002
C-1_2026-10-06-11-04-57	2026-10-06	16	8	45.9	1015
C-2_2026-10-06-11-10-33	2026-10-06	21	9	50.3	1015
C-3_2026-10-06-11-20-51	2026-10-06	16	8	43.8	1015
C-4_2026-10-06-11-45-54	2026-10-06	12	6	32.8	1015
C-5_2026-10-06-11-33-51	2026-10-06	12	6	32.0	1015
C-6_2026-10-06-11-58-21	2026-10-06	22	11	63.1	1015
C-7_2026-10-06-12-09-56	2026-10-06	17	9	48.7	1015
C-8_2026-10-06-12-15-48	2026-10-06	17	9	47.8	1015
D-1_2026-09-29-17-08-46	2026-09-29	8	3	16.4	1002
D-2_2026-09-29-17-15-57	2026-09-29	13	5	31.3	1002
D-3_2026-09-29-17-20-32	2026-09-29	18	8	42.9	1015
D-4_2026-09-29-17-33-25	2026-09-29	17	8	43.1	1002
D-5_2026-09-29-17-44-05	2026-09-29	8	3	16.6	1002
E-1_2026-10-06-15-05-45	2026-10-06	7	4	16.7	1015
E-2_2026-10-06-15-09-30	2026-10-06	10	5	24.3	1015
E-3_2026-10-06-15-13-04	2026-10-06	12	5	30.3	1015
E-4_2026-10-06-15-17-20	2026-10-06	11	5	27.1	1015
E-5_2026-10-06-15-21-16	2026-10-06	12	6	29.5	1015
E-6_2026-10-06-15-25-25	2026-10-06	6	3	14.2	1015
F-1_2026-10-06-15-36-52	2026-10-06	20	10	49.5	1015
F-2_2026-10-06-15-51-37	2026-10-06	16	8	37.9	1015
F-3_2026-10-06-16-05-27	2026-10-06	5	2	11.6	1015
F-4_2026-10-06-16-10-18	2026-10-06	19	8	44.6	1015
F-5_2026-10-06-16-18-39	2026-10-06	17	7	40.3	1015

Maintainers: making the next release

Releases are made with scripts/hf_release.py of the parser repository (ROSbag_nuscenes_parser, branch tcar-gui; the commit of each release is in the table above):

# new recordings: parse them, curate them in the TCAR Parser app, press 최종 확정, then
python scripts/hf_release.py --branch <new> --from 1015 --repo shchon11/TCar_hyu --dataroot <full dataroot> --token-file <hf token file>            # plan
python scripts/hf_release.py --branch <new> --from 1015 --repo shchon11/TCar_hyu --dataroot <full dataroot> --token-file <hf token file> --upload   # build + upload

It refuses to publish unless every scene is locked and the dataset passes the checks (table references, devkit, CAN bus), and unless every tar already published still matches the dataset. Only the new recordings' tars are built and uploaded (one commit each, resumable); the tables, manifest, this card and assemble.py go last. First release of this series: 1002.