Skip to content

Datasets

All benchmark data lives in the Calibench Hugging Face organization. The calibench-vi/datasets repository is the source of truth: it hosts our own VI-Rig captures (CC-BY-4.0) and ships download recipes for the third-party datasets the benchmark evaluates against. The Google Drive links from the legacy site are retired.

Collected on a self-built VI-Rig for research on multi-visual-inertial calibration. Reference Kalibr and MVIS results for these captures are published in calibench-vi/results.

DatasetSensorsSequencesData
4 IMUs + 3 cameras4 IMUs, 3 global-shutter10 AprilGridvirig-4imus-3cams
4 IMUs + 3 cameras — ArUcosame rig, ArUco-tag sequences17virig-4imus-3cams-aruco
4 IMUs + 4 cameras4 IMUs, 2 GS + 2 rolling shutter15 AprilGrid (+7 raw)virig-4imus-4cams
5 IMUs + 4 cameras5 IMUs (two T265 stereo fisheye pairs), 4 cameras7 AprilGridvirig-5imus-4cams

Calibench hosts only data we own; every third-party dataset the benchmark evaluates ships as a download.py recipe (manifest + checksums) that pulls it from its original host — EuRoC MAV, TUM VI, TUM RGB-D, TUM VIE, TUM Double-Sphere, TUM RS, Monado MSD, Hilti 2022, Hilti 2023 Handheld, PennCOSYVIO, and an urban fisheye set. See recipes/ and its SOURCES.md for the exact list and licenses. The leaderboard evaluates calibration methods across these datasets plus the hosted VI-Rig captures.

Datasets with complete MVIS runs documented as worked examples on this site:

DatasetRigExample
TUM VIStereo 1024² fisheye + BMI160TUM VI
TUM MIC (Monado / Valve Index)Stereo fisheye + IMUMonado
looperrobotics SA162–3 cameras + 400 Hz IMULooper SA16

Contributing one: Submit a dataset.

If you use the Calibench data, please cite the benchmark, and the MVIS paper for the VI-Rig captures:

@misc{calibench2026,
title = {Calibench: Calibration Benchmark Datasets},
author = {Yang, Yulin},
year = {2026},
howpublished = {\url{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},
doi = {10.1177/02783649241245726}
}

Third-party datasets keep their own licenses and citations — the dataset card maps each config to its required upstream citation.

We use a self-built VI-Rig to collect data. The rig contains 4 IMUs, one monocular camera, and 2 stereo pairs. Treating each stereo camera as two monocular cameras, we have 5 cameras total.

VI-Rig with 4 IMUs and 5 cameras

Visual-Inertial (VI)-Rig — 4 IMUs and 5 cameras:

  • MicroStrain GX3-25
  • MicroStrain GX3-35
  • Xsens MTI-100
  • RealSense T265 IMU
  • FLIR BlackFly
  • RealSense T265 Left & Right cameras
  • ELP Left & Right rolling-shutter cameras

ROS drivers can be found here.

The datasets contain the following ROS topics:

SensorHzTopicResolution
MicroStrain GX3-25500/gx3_25/data
MicroStrain GX3-35100/gx3_35/imu/data
Xsens MTI-100400/imu/data
RealSense T265 IMU200/t265/imu
FLIR BlackFly30/blackfly/image_raw720×540
RealSense T265 Left30/t265/fisheye1/image_raw848×800
RealSense T265 Right30/t265/fisheye2/image_raw848×800
ELP Left rolling shutter25/elp/split_sync_image_node/left/image_raw640×480
ELP Right rolling shutter25/elp/split_sync_image_node/right/image_raw640×480