Comparison vs Kalibr / Basalt

Use this page to see where MVIS sits next to Kalibr and Basalt, the two open-source camera–IMU toolboxes most people use. Kalibr is mature and widely validated, and MVIS does not aim to replace it where it works well.

Checked against: MVIS 0.1.0 and our benchmark runs of 10 November 2026 (Methodology).
The numbers: per-dataset results of all three tools are on the Leaderboard.

Feature matrix

Capability	MVIS	Kalibr	Basalt
Camera intrinsics + distortion	✅	✅	✅
Distortion models	radtan, kb4, equidistant, FOV, DS, EUCM, UCM, omni_radtan (experimental)	radtan, equidistant, FOV, omni, omni-radtan, DS, EUCM	radtan, KB4, DS, EUCM, UCM
Different models in one rig	✅	✅	✅
Camera–IMU extrinsics + time offset	✅	✅	✅
Multiple IMUs in one solve	✅	✅	✖ one IMU per run
Gyro-only mode per auxiliary IMU	✅	✖	✖
IMU intrinsics (scale, misalignment)	✅ — 3 conventions (`kalibr`, `rpng`, `calyx`)	✅	✅ scale + misalignment
G-sensitivity `Tg`	✅	✅	✖
Gyro↔accel offset `t_ga`	✅ on by default	✖	✖
Board shape (one height per corner)	✅ `estimate.board_points`	✖ flat board	✖ flat board
KB4-type polynomial past 90°	✅ `kb4`, what `fisheye` means	✖ pinhole-equi needs rays in front of the camera	✅ KB4
Rolling-shutter readout	✅ in camera and camera–IMU solves	camera-only, separate tool	✖
Global + rolling shutter + IMUs in one solve	✅ readout per camera	✖	✖
Corrected-dataset export	✅ `apply-calib`	✖	✖
Digital-twin simulation	✅ `mvis.simulation`	✖	✖
Round-trip PASS/FAIL validation	✅ `roundtrip`	✖	✖
Batch sweep over a dataset family	✅ `input.bags`	✖	✖
Dataset download + check	✅ `mvis datasets`	✖	✖
Trajectory model	discrete factor graph	continuous-time B-spline	continuous-time B-spline
Boards	AprilGrid, chessboard, ArUco	AprilGrid, chessboard, circles	AprilGrid
Input	ROS1/ROS2 bag, EuRoC, image folders	ROS bag	ROS bag, EuRoC
Runs without ROS	✅ pure Python	✖	✅ C++ build, no ROS
Platforms	Linux, macOS (Python 3.11–3.12)	Linux (ROS 1, or Docker)	Linux, macOS
Kalibr interop	✅ read + write camchain, read AprilGrid targets	native	✖

What is different

No ROS, no C++ build. MVIS installs with pip and reads ROS1 and ROS2 bags itself. Kalibr needs a ROS workspace; Basalt needs a C++ build. Requirements.
More than one IMU, each with its own mode. MVIS puts every IMU in one graph, and each auxiliary IMU (every IMU but the base) can run gyro-only when its accelerometer is unreliable. Kalibr also solves several IMUs in one run. Basalt takes one IMU per run, so a four-IMU rig is calibrated IMU by IMU and composed.
It fits the gyro–accel offset. t_ga is the time offset between the gyroscope and accelerometer channels of one IMU. It is on by default in camera–IMU runs. Kalibr and Basalt give an IMU one clock.
- On two real IMUs (the Looper Insight 9 and the T265) MVIS measures a t_ga of about −5 to −6 ms.
- There, Kalibr’s and Basalt’s camera time offset td differs from MVIS by about 1.6–3.2 ms. A sweep in which we delay only the accel predicts most of that gap from t_ga alone.
- Not yet tested: which td is closer to the truth. Details.
It can solve the board shape. With estimate.board_points: true MVIS solves one height per corner. Kalibr and Basalt take the board as flat. On our VI-Rig mount a flat solve shifts the focal length, gyro scale and lever. Details.
Rolling shutter in the camera–IMU solve. MVIS solves the readout time together with the IMU. Kalibr’s rolling-shutter tool is camera-only.
Validation is a command. mvis roundtrip applies the calibration to its own recording, calibrates again and prints PASS or FAIL. A PASS needs the IMU intrinsics to come back as identity and the geometry to stay put. Neither Kalibr nor Basalt ships a check like it. Round-trip validation.
Discrete-time. MVIS places pose, velocity and bias states at a fixed rate (state_hz, 30 Hz by default). Kalibr and Basalt fit a B-spline, a smooth curve in continuous time.