Skip to content

Overview

MVIS is open-source multi-visual-inertial sensor calibration: a Python toolbox for camera and camera–IMU calibration, from a single camera to multi-camera, multi-IMU rigs. An IMU (inertial measurement unit) is the sensor with a gyroscope and an accelerometer.

  • Package: pymvis. You import it as mvis and run it as the mvis command.
  • Runs on: Linux and macOS (Windows tentative), Python 3.11 and 3.12, with no ROS install. See Requirements.
  • Code: github.com/openmvis/openmvis.

OpenMVIS, the project around MVIS, also publishes calibration datasets, a benchmark against Kalibr and Basalt, and a device wiki of calibrated sensors.

MVIS solves the quantities below from one recording, in a camera solve and then a camera–IMU solve.

QuantityScopeNotes
Camera intrinsics + distortionper cameraFocal length, image centre and lens model. Eight models, mixable in one rig (Camera model)
Camera extrinsicsper cameraWhere each camera sits relative to the reference camera
Camera–camera time offsetper cameraFor cameras that are not hardware-synced
Rolling-shutter readoutper rolling-shutter cameraThe time the camera takes to read its rows, top to bottom (why)
Board poses and board shapeper frame / per cornerThe shape is one height per corner, off by default (why)
Camera–IMU extrinsics T_CtoI and time offset tdper camera
IMU–IMU extrinsics and time offsetper auxiliary IMUEach extra IMU relative to the base IMU, the first IMU in the config
IMU intrinsics Dw, Da, optional Tgper IMUGyro and accel scale and misalignment, in one of three conventions (IMU model)
IMU biasesper IMU
Gyro–accel time offset t_gaper IMU (not gyro-only IMUs)The delay between the gyro and accel channels of one IMU, on by default (why)

Several of these have no counterpart in Kalibr or Basalt. The comparison lists them.

A calibration has two solves and a check.

  1. Cameras: mvis cam solves each camera’s lens and the camera–camera extrinsics from a board recording. It writes a camera YAML and an HTML report.
  2. Cameras + IMUs: mvis cam-imu reads that camera YAML as its starting point and solves the camera–IMU extrinsics, time offsets and IMU intrinsics. It writes a JSON result, a calibration YAML, a Kalibr camchain where the camera model allows, and an HTML report.
  3. Check: mvis verdict grades a result GREEN, YELLOW or RED. mvis roundtrip applies a camera–IMU calibration to its own recording and calibrates again: the IMU intrinsics must come back as identity and the geometry must not move. A digital twin of your rig tells you what the recording can pin down before you trust it.

mvis run starts either solve from the config’s mode: field. Every command is in the CLI reference.

Inside the camera solve, the steps are:

  1. Detection: find the board corners in every image.
  2. Quality gating: drop frames with too few tags and corners seen too rarely.
  3. Init: a first guess of each lens and each board pose.
  4. Per-camera solve: refine each camera on its own, with its rolling-shutter readout if it has one.
  5. Alignment: find where the cameras sit relative to each other, and their time offsets.
  6. Joint solve: refine all cameras together.
  7. Outlier rerun: drop corners with large error and solve again.

The camera–IMU solve starts from the camera YAML, picks the part of the recording with the most motion and runs one joint solve with the IMU data.

You can also start from an existing calibration, such as Kalibr output, and pin what you trust. See Configuration.