Datasets

This page lists every dataset we use to test and benchmark MVIS, where to get it, and its licence.

• Where the data lives: the calibench-vi/datasets repository on Hugging Face. Calibench is the name of our calibration benchmark.
• What it hosts: our own VI-Rig recordings, two partner recordings, third-party sets whose licence allows us to share them again (CC BY 4.0), and result files for some sets (tum-fisheye/results/, uzh-fpv/results/, virig-5imus-4cams/results/).
• What it does not host: all other third-party sets. For those, a download recipe fetches the files from the original site.
• Results: the reference results for each set are on the leaderboard and in calibench-vi/results.

VI-Rig: our own recordings

We built the VI-Rig to research calibration of rigs with several cameras and several IMUs. An IMU (inertial measurement unit) is a gyroscope plus an accelerometer. Each dataset below is one sensor setup on the same frame, recorded as ROS1 bags (one bag per recording) while we move the rig in front of a calibration board.

Dataset	Cameras	IMUs	Board	Recordings	Size	Results
4 IMUs + 3 cameras	Blackfly + T265 pair, all global shutter	4, 100–500 Hz	AprilGrid	10, 131–161 s	76 GB	results · example
4 IMUs + 3 cameras, ArUco	same 3 cameras	4 IMUs (3 in recordings 11–17)	ArUco tags	17, 102–139 s	117 GB	not benchmarked
4 IMUs + 4 cameras	T265 pair (global) + ELP stereo (rolling)	4, 100–500 Hz	AprilGrid	15, 131–152 s	113 GB	results · example
5 IMUs + 4 cameras	two T265 pairs, all global shutter	5, 100–500 Hz	AprilGrid	7, 41–64 s	28 GB	results · example

Terms used in the table:

• Global shutter: the camera exposes every row of the image at the same time.
• Rolling shutter: the camera reads the image out row by row, so each row has its own capture time.
• AprilGrid: a printed grid of AprilTags, square codes that the software finds and identifies in each image.
• ArUco: another family of square marker tags.

Licence and access

• Licence: CC BY-NC 4.0. Free for research, teaching and other non-commercial use, with attribution.
• Gated: log in to Hugging Face and accept the terms on the dataset page. Access is granted at once.
• Benchmark exception: anyone, a company included, may run a calibration tool on these recordings to produce Calibench results and publish them with attribution.
• Other commercial use: needs a separate licence. Ask in the Community tab of the Hub repository or at licensing@openmvis.com.

Download

Accept the terms on the Hub first, then set HF_TOKEN to a read token from your Hugging Face settings (or log in once with hf auth login, after pip install huggingface_hub).

mvis datasets list                                          # every dataset key, recording and size
mvis datasets download --select mvis_4imus_3cams/april_01   # one recording, 6.9 GB

• Where files land: in $MVIS_DATASETS (default ~/Datasets). The example configs refer to the data as ${MVIS_DATASETS}/....
• Checks: mvis datasets checks each file’s size and SHA-256 checksum.
• Without MVIS: use the Hugging Face CLI; each dataset page gives the exact command.

Partner recordings on the Hub

Two partners agreed to share their recordings under the same terms as the VI-Rig sets (CC BY-NC 4.0, gated, with the benchmark exception).

Dataset	Rig	Recordings	Hub folder	On this site
Looper Robotics Insight 9	2 global-shutter mono cameras + 1 rolling-shutter colour camera + 400 Hz IMU	10 ROS1 bags, AprilGrid	looper-insight9	device · results · example
GiLabs examples	GI Ego1 and GI EGO1GS head-worn cameras: 2 cameras + IMU	3 MCAP recordings, not in the benchmark	gilabs-examples	GI Ego1 · GI EGO1GS

Third-party datasets

These sets come from other groups. Each keeps its own licence and citation; please cite the original authors.

• Hosted: the calibration recordings of sets under CC BY 4.0, converted to ROS1 bags. A .provenance.json beside each bag records the source address, the source checksum and the conversion steps.
• Recipe: we never copy these files. The Hub’s recipes/download.py fetches them from the original site and checks them.
• Key: mvis datasets download --dataset <key> also fetches from the original site and checks each file. For the hosted sets it gets the original archives; mvis datasets convert then turns them into ROS1 bags. Urban fisheye and part of PennCOSYVIO have no checksum in the catalog; add --allow-unverified to fetch them unchecked.

Dataset	Rig	Licence	How to get it	Original source	On this site
TUM-VI	stereo fisheye 1024² (and 512²) + BMI160 IMU	CC BY 4.0	recipe · keys tum_vi_raw, tum_vi_standard_512	TUM	results · example
Monado SLAM, MIC (Valve Index)	2 fisheye + 1 kHz IMU	CC BY 4.0	hosted: converted/tum_mic · key tum_mic	Collabora	device
Monado SLAM, MGC (HP Reverb G2)	4 fisheye + 1 kHz IMU	CC BY 4.0	hosted: converted/monado_mgc · key monado_mgc	Collabora	device · results · example
Monado SLAM, MOC (Samsung Odyssey+)	2 fisheye + 1 kHz IMU	CC BY 4.0	hosted: converted/monado_moc · key monado_moc	Collabora	device
TUM-VIE, frame cameras	stereo 1024² + BMI160 IMU	CC BY 4.0	hosted: converted/tum_event · key tum_event	TUM	–
UZH-FPV	DAVIS346 + IMU; Snapdragon stereo + IMU	CC BY-NC-SA 3.0	keys uzh_fpv_davis, uzh_fpv_snapdragon	UZH	results · example
TUM-Fisheye (Double Sphere)	1 camera, 5 fisheye lenses	none stated	recipe · key tum_fisheye	TUM	Why kb4 · MVIS results on the Hub
TUM RGB-D	Kinect and Xtion RGB + IR cameras	CC BY 4.0	recipe · key tum_rgbd_bag	TUM	–
PennCOSYVIO	GoPro, Tango and VI-Sensor cameras + IMUs	citation only	recipe · key penncosyvio	Penn	–
Urban fisheye (ImprovedOcamCalib)	fisheye and GoPro stills	GPL v2	recipe · key fisheye_dataset	GitHub	–
EuRoC MAV	stereo + IMU	CC BY	Hub recipe only	ETH Zurich	–
TUM RS (rolling shutter)	rolling- and global-shutter cameras + IMU	none stated	Hub recipe only	TUM	–
Hilti 2022	multi-camera + IMU	CC BY-NC-SA 3.0	Hub recipe only	Hilti	–
Hilti 2023 handheld	multi-camera + IMU	CC BY-NC-SA 3.0	Hub recipe only	Hilti	–

• Exact sources and licences: see recipes/SOURCES.md on the Hub.
• Add your own: see Submit a dataset.

Citation

If you use the Calibench data, please cite:

• Calibench, for any of the data.
• The MVIS paper, for the VI-Rig sets with 4 IMUs (3 cameras, 3 cameras ArUco, 4 cameras). The 5-IMU set has no paper yet; cite Calibench for it.
• The original paper of every third-party set you use. The dataset card lists them.

The MVIS paper entry below is the same as on Contributors & citation, which also has the entries for the software and the benchmark results.

@misc{calibench2026,
  author       = {Yang, Yulin},
  title        = {Calibench: Visual-Inertial Datasets for Calibration Benchmark},
  year         = {2026},
  howpublished = {Hugging Face Datasets},
  note         = {https://huggingface.co/datasets/calibench-vi/datasets}
}

@article{yang2024mvis,
  title   = {Multi-visual-inertial system: Analysis, calibration, and estimation},
  author  = {Yang, Yulin and Geneva, Patrick and Huang, Guoquan},
  journal = {The International Journal of Robotics Research},
  year    = {2024},
  volume  = {43},
  number  = {13},
  pages   = {1995--2026},
  doi     = {10.1177/02783649241245726}
}