UoSM-Campus — handheld LiDAR-visual-inertial dataset

This repository hosts a handheld multi-sensor dataset collected around the University of Southampton Malaysia campus. It contains nine sequences recorded with a Livox Mid-360, a ZED 2i stereo camera, and a HikRobot camera. The data is intended for evaluating LiDAR-visual-inertial odometry, SLAM, calibration, and sensor-fusion methods.

The release includes raw ROS 2 MCAP recordings, ZED SVO2 files, calibration assets, and pseudo ground-truth trajectories. The total trajectory length is about 1.9 km.

Recorded with the handheld mapper: https://github.com/limshoonkit/handheld_mapper

---

Dataset summary

• Nine handheld sequences
• Day and night recordings
• Stereo images, IMU, LiDAR, and pseudo ground truth
• Suitable for LiDAR-visual-inertial odometry, calibration, and benchmarking

Data layout

1_UoSM-Campus/
├── <SEQ>/
│   ├── rosbag_<timestamp>/          ROS 2 MCAP — all sensors, host clock
│   └── svo_<timestamp>/             ZED .svo2 — camera-native clock
├── LF-02-N/
│   └── rosbag_svo_reexport/         see "LF-02-N" below
├── ground_truth/                    TUM format (.txt)
│   ├── fastlivo2/<SEQ>.txt          FAST-LIVO2 odometry only
│   └── fastlivo2_lvba_blended/<SEQ>.txt   FAST-LIVO2 + Global-LVBA (use this)
├── FastCalib-HikCam_Mid360/         HikRobot camera <-> Mid-360 extrinsics
└── kalibr_SLZed_HikCam/             ZED <-> HikRobot Kalibr calibration

Sequences

BF = both floor, LF = lower floor, UF = upper floor; -D = day, -N = night.

Path is distance travelled; Span is the diagonal of the trajectory's axis-aligned bounding box, i.e. how far the sequence spreads spatially. A sequence can have a large path but small span.

Sequence	Poses	Duration [s]	Path [m]	Span [m]
BF-00	2898	416.3	507.4	89.9
LF-01-D	1240	168.0	184.3	48.4
LF-01-N	1397	189.3	183.5	54.6
LF-02-D	1037	145.1	158.3	64.1
LF-02-N	912	129.8	141.6	57.8
UF-01-D	999	140.7	162.9	67.2
UF-01-N	965	137.4	152.0	66.0
UF-02-D	1217	175.9	204.1	65.2
UF-02-N	1228	180.2	198.4	69.0

LF-03-D is not part of this release. A roughly 520 ms backward host system-clock step at about 58 s corrupted the ZED IMU, the Livox IMU, and the image stamps together. The same step also affected the Livox stream that the pseudo ground truth is built from, so the reference is degraded there as well (600 ms pose gap, 55.6 cm jump at about 56.8 s). It is not a scoreable datapoint, and its .svo2 has its own 5.26 s IMU hole, so a re-export cannot recover it.

Sensor topics

Topic	Type	Rate
/zed_node/left/color/rect/image	sensor_msgs/msg/Image	30 Hz
/zed_node/right/color/rect/image	sensor_msgs/msg/Image	30 Hz
/zed_node/imu/data	sensor_msgs/msg/Imu	202 Hz
/livox/lidar	sensor_msgs/msg/PointCloud2	10 Hz
/livox/imu	sensor_msgs/msg/Imu	200 Hz
/left_camera/image	sensor_msgs/msg/Image	10 Hz

---

Ground truth (TUM format)

Both sets are plain-text TUM trajectory format: one pose per line, space separated, no header.

timestamp tx ty tz qx qy qz qw
1590192263.400343 -0.035121 0.080318 0.018658 0.007820 0.222244 -0.002448 0.974957

Translation is in metres, and rotation is given as a unit quaternion $(x, y, z, w)$ with $w$ last. Timestamps are on the LiDAR clock, one pose per LiDAR scan (about 10 Hz). The files are directly loadable by evo and by any tool that reads TUM format.

• ground_truth/fastlivo2/ — raw FAST-LIVO2 LiDAR-visual-inertial odometry. Locally accurate, but it drifts over a full loop.
• ground_truth/fastlivo2_lvba_blended/ — FAST-LIVO2 refined by Global-LVBA (global LiDAR-visual bundle adjustment) and blended back onto the FAST-LIVO2 timeline. This is the reference to evaluate against. It is a pseudo ground truth: there is no external motion capture or RTK, so treat it as accurate to roughly the centimetre rather than the millimetre.

The blended file has slightly fewer poses than the FAST-LIVO2 one (for example, 912 versus 925 on LF-02-N), because the ends are trimmed where the BA window has no support.

---

LF-02-N — the ZED dropout and how to re-export

The original MCAP for LF-02-N is missing about 2.86 s of the entire ZED stream around $t \approx 53$ s. Both images and the ZED IMU stop together while the Livox stream continues. The maximum image gap is about 2860 ms against a nominal 33 ms interval, so any VIO run on this bag has to coast blind through nearly three seconds.

Re-exporting the ZED stream from the .svo2 recovers the missing data. The rosbag_svo_reexport/ directory contains that re-export. It carries only the three ZED topics, which is the complete stereo-inertial VIO input. The Livox stream and therefore the ground truth still come from the original MCAP.

How to re-export

The export opens the .svo2 with the ZED SDK, walks it frame by frame, and writes /zed_node/{left,right}/color/rect/image plus /zed_node/imu/data into an MCAP at SDK timestamps. Two details are essential:

1. The coordinate system must be COORDINATE_SYSTEM.RIGHT_HANDED_Z_UP_X_FWD. The SDK default is IMAGE, which places gravity on $-Y$ instead of $+Z$. RIGHT_HANDED_Z_UP fixes gravity to $+Z$ but is still yawed 90 degrees from ROS, so acceleration and gyroscope $x/y$ values come out swapped and negated. The IMU rate is also around 400 Hz rather than 200 Hz.
2. Angular velocity needs deg2rad. The SDK reports degrees per second, while the ROS wrapper publishes radians per second.

---

Benchmark results

ATE (APE RMSE, m) against ground_truth/fastlivo2_lvba_blended/, SE(3)-Umeyama aligned, per-sequence clock-offset fitted by speed-profile cross-correlation. Every cell below is a completed run with no divergences or crashes.

DL-VINS front-ends from https://github.com/limshoonkit/DL-VINS-Factory-ROS2 were run stereo-inertial on the ZED camera stream. The scored stream is /dl_vins/odometry (no loop closure), so these are pure odometry numbers. ZED GEN_3 is the ZED SDK's own onboard VIO at GEN_3 positional-tracking mode for reference.

Method	BF-00	LF-01-D	LF-01-N	LF-02-D	LF-02-N	UF-01-D	UF-01-N	UF-02-D	UF-02-N
ZED GEN_3	1.43	0.38	0.51	0.69	0.84	2.38	0.72	2.28	0.47
ALIKED + LightGlue	1.37	0.49	0.35	0.73	0.57	1.01	0.94	0.99	0.58
ALIKED + LK	1.12	0.62	0.69	0.39	0.98	1.01	3.28	1.76	1.57
SuperPoint + LightGlue	1.82	0.59	0.30	0.99	0.42	1.09	1.48	1.12	0.69
SuperPoint + LK	1.15	0.40	0.93	0.83	0.64	1.26	2.55	2.24	2.09
XFeat + LightGlue	1.33	0.51	0.88	0.95	0.98	2.08	3.29	1.66	1.16
XFeat + LK	1.71	0.54	0.92	0.59	1.05	1.03	2.68	2.03	2.53
RaCo + LightGlue	2.22	0.46	0.40	0.39	0.66	1.12	0.63	0.94	0.64
RaCo + LK	1.30	0.57	0.79	0.37	1.14	1.38	2.79	2.14	1.09
GFTT (classical)	1.61	0.53	0.43	0.53	0.78	0.83	2.34	1.83	0.78

Summary over all 9 sequences:

Method	median	mean	worst	x	y	z
RaCo + LightGlue	0.64	0.83	2.22	0.62	0.49	0.15
ZED GEN_3	0.72	1.08	2.38	0.64	0.60	0.54
ALIKED + LightGlue	0.73	0.78	1.37	0.59	0.44	0.17
GFTT	0.78	1.07	2.34	0.77	0.66	0.19
SuperPoint + LightGlue	0.99	0.94	1.82	0.76	0.51	0.15
ALIKED + LK	1.01	1.27	3.28	0.97	0.71	0.25
XFeat + LK	1.05	1.45	2.68	1.11	0.82	0.28
RaCo + LK	1.14	1.29	2.79	0.95	0.74	0.25
SuperPoint + LK	1.15	1.34	2.55	1.02	0.76	0.26
XFeat + LightGlue	1.16	1.43	3.29	1.09	0.82	0.16

---

Known issues

• LF-02-N: 2.86 s ZED dropout in the MCAP. Use rosbag_svo_reexport/ for any ZED-based VIO. The Livox stream in the original bag is unaffected.
• LF-03-D: excluded from the release.
• The MCAP record-clock epoch is per recording session, not global. Do not reuse a time offset fitted on one sequence for another. Values are tabulated below.

Clock offsets

The bag carries two independent clocks. The ZED topics are stamped on the host clock, while /livox/* is stamped on the Livox clock, and the ground truth is built from the Livox stream. Therefore the ground truth is on the Livox clock. To compare a ZED-based trajectory against the ground truth, subtract $\Delta$:

t_livox = t_host - Δ

Sequence	Δ [s]	Recording session
BF-00	176718469.580	2025-12-28 (BF)
LF-01-D	176708459.435	2025-12-28 (day field)
LF-02-D	176708459.426	2025-12-28 (day field)
UF-01-D	176708459.412	2025-12-28 (day field)
UF-02-D	176708459.416	2025-12-28 (day field)
LF-01-N	176650851.002	2025-12-27 (LF night)
UF-01-N	176648552.016	2025-12-27 (UF night)
UF-02-N	176648552.008	2025-12-27 (UF night)
LF-02-N	178551189.897	2026-01-18

Derived as median(log_time - header.stamp) over /livox/imu minus the same over /zed_node/imu/data, so the shared record latency cancels and what remains is the epoch difference. Measured over an 8 s window at each end of every bag, $\Delta$ is constant to within about $\pm 4$ ms (drift $< 30 ; \mu s/s$), so a single constant per sequence is sufficient.

The four day sequences share one session and agree to 23 ms; the two UF night sequences agree to 8 ms; LF-02-N, recorded three weeks later, is 1.84e6 s away. That is why a single global constant cannot work.

The LF-02-N re-export is on a third clock — the ZED SDK camera clock rather than the host clock — so the $\Delta$ above does not apply to rosbag_svo_reexport/. Fit that one directly; the offset is constant and the drift is less than $39 ; \mu s/s$ over 128 s, so most evaluation pipelines that cross-correlate speed profiles will pick it up automatically. The original LF-02-N recording from 2025-12-27 had to be discarded and re-recorded three weeks later because that sequence contains many humans walking around the scene, which corrupts the overall ground truth.

---

License and usage notes

Please treat this release as a research dataset with a pseudo ground truth. Use the data responsibly and cite the repository if you build on it.