Calibrate a whole rig at once

From one recording to a full sensor model.

MVIS is the calibration library behind the OpenMVIS benchmark. It solves camera intrinsics and distortion, extrinsics, time offsets, rolling-shutter readout, board poses and (optionally) the board shape, camera–IMU transforms, IMU intrinsics and biases in a single GTSAM factor graph.

It reads ROS1 and ROS2 bags through rosbags, so ROS does not need to be installed. It also reads EuRoC/ASL folders, image folders and timestamped images.

# cameras only
$ mvis cam -c camera.yaml

# cameras + IMUs, jointly
$ mvis cam-imu -c cam_imu.yaml

# apply the result and recalibrate; PASS / FAIL
$ mvis roundtrip -c cam_imu.yaml

One self-contained HTML report per run.

Not in Kalibr or Basalt

Six things the common tools do not do.

Kalibr and Basalt are the open-source camera–IMU toolboxes most people use. Neither covers these six cases. See the full comparison.

1. Rolling shutter + IMU

   A rolling-shutter camera reads its rows one after another. MVIS solves that readout time together with the IMU. Kalibr's rolling-shutter tool is camera-only; Basalt has none.

   GuideWhy it matters

2. Global + rolling shutter + IMU

   Global-shutter and rolling-shutter cameras on one rig, solved with all its IMUs in one run. Our VI-Rig recording has 2 global-shutter cameras, 2 rolling-shutter cameras and 4 IMUs.

   ExampleConfig

3. Board shape and pose

   MVIS solves the board pose and, optionally, one height per corner (estimate.board_points). Kalibr and Basalt assume a flat board. On our VI-Rig mount, a flat solve biases the focal length by up to 1 % and the gyro scale by about 1 %.

   Why it matters

4. Gyro-to-accel time offset

   t_ga is the time offset between an IMU's gyroscope and its accelerometer. MVIS fits it by default in camera–IMU runs. Kalibr and Basalt use one IMU clock, so the offset leaks into the camera time offset.

   Why it matters

5. Digital-twin simulation

   A digital twin is a simulated copy of your recording where the truth is known. MVIS re-solves it and reports the error of each parameter, so you can test the solver on your own rig.

   Guide

6. Round-trip verification

   mvis roundtrip applies the calibration to its own data and calibrates again. It reports PASS or FAIL without ground truth, and exits non-zero on FAIL so it can run in CI.

   Guide

What the graph estimates

The state, not a sequence of stages.

Quantity	Symbol	Per	Unit	Notes
Intrinsics + distortion	f, c, k	camera	px	Eight models, mixable within one rig
Camera–IMU extrinsics	T_CtoI	camera	m · log(R) rad	Every camera to the base IMU, the reference IMU of the rig
Camera time offset	t_d	camera	ms	About −123 ms of USB lag on the VI-Rig ELP stereo
Rolling-shutter readout	t_r	camera	ms	Each corner projected at its own row time
IMU–IMU extrinsics	T_IstoI	aux IMU	m · log(R) rad	Plus a per-IMU time offset
IMU intrinsics	D_w, D_a	IMU	—	kalibr, rpng or calyx convention
Gyro–accel offset	t_ga	IMU	ms	About −5.7 ms on the T265 IMU; why it matters
Biases	b_g, b_a	IMU	rad/s · m/s²	Along the whole trajectory
Board pose + trajectory	T_BinW, T_ItoW(t)	recording	m · log(R) rad	Chessboard, AprilGrid or ArUco targets
Board shape	z_k	corner	mm	Optional, off by default; on for our VI-Rig runs; why it matters

Camera models

Eight projection models.

• radtanPinhole, radial–tangential
• kb4Kannala–Brandt fisheye, the default; why
• equidistantKB polynomial, under 90°
• fovField-of-view model
• double_sphereWide-angle fisheye
• eucmExtended unified camera
• ucmUnified camera
• omni_radtanOmnidirectional

Benchmark

Checked against Kalibr on TUM-VI.

0.2px

largest median focal-length (fx, fy) difference from Kalibr; principal point within 0.35 px. Camera-only stereo, 8 raw recordings, view angles cut to 85° or less for every tool

<0.1mm

median stereo-baseline difference, inside the 0.1 mm run-to-run spread; same 8 recordings

1.8mm

largest median lever-arm difference (camera position in the IMU frame), 4 raw and 4 512×512 recordings. Kalibr was given MVIS's intrinsics and baseline here

0.9ms

median camera time offset above Kalibr's; Basalt is within 0.1 ms of Kalibr. MVIS also fits t_ga (−1.5 to −1.6 ms here), which Kalibr does not have

Benchmark batch 2026-11-10a. See the Leaderboard and each device's calibration.

VI-Rig, our own rig: up to 5 IMUs and 4 cameras across its recordings. They are in Datasets.

It checks its own work

A low residual is not the test.

A low reprojection error says a calibration fits its own data, not that it is right. In a round trip, PASS needs the IMU intrinsics and biases to collapse to identity and zero, and the geometry to hold. See Round-trip validation.

Inputs

ROS1 and ROS2 bags, EuRoC/ASL folders, image folders, timestamped images

Targets

AprilGrid, chessboard, ArUco

Interop

Reads and writes Kalibr camchain and camchain-imucam

Export

mvis apply-calib writes a copy of a raw ROS1 bag with the calibration applied

Cite

If you use the toolbox or the benchmark.

Benchmark

@misc{openmvis,
  title  = {{OpenMVIS}: A Benchmark for Multi-Visual-Inertial Sensor Calibration},
  author = {{OpenMVIS Contributors}},
  year   = {2026},
  url    = {https://openmvis.com}
}

Method

@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}
}

Cite leaderboard results as OpenMVIS leaderboard, batch 2026-11-10a.

Community

Bring your rig.

OpenMVIS is open: the toolbox, the benchmark records and our own VI-Rig recordings. Add your device, your dataset or your results, or tell us where MVIS breaks.

• Submit a dataset
• Submit results
• Device wiki
• GitHub

Visitors to this page.