TUM VI

The TUM Visual-Inertial dataset is the reference check for a camera–IMU calibrator: a stereo-fisheye + IMU rig with a published Kalibr calibration to score against. It is also the dataset where every distortion model can be compared on identical data.

The rig

Sensor	Rate	Detail
2 × monochrome global-shutter cameras	20 Hz	1024×1024 fisheye
Bosch BMI160 IMU	200 Hz	/imu0

Board: AprilGrid 6×6, tag_size 0.088 m, tag_spacing 0.3, family t36h11.

Getting the data

Use the raw IMU-calibration bags — dataset-calib-imu{1,2,3,4}.bag — for the camera–IMU solve, and standard_1024/dataset-calib-cam1_1024_16.bag as the board bag when a model needs a camera-only fit first.

tum_vi_calib/
  raw/dataset-calib-imu{1..4}.bag             # cam-IMU sequences
  raw/dataset-calib-cam1-camchain.yaml        # Kalibr omni/radtan intrinsics
  tumvi_camchain.yaml                         # Kalibr equidistant intrinsics
  standard_1024/dataset-calib-cam1_1024_16.bag
  omni-radtan-1024/ , pinhole-equi-1024/      # Kalibr cam-IMU reference results

One step, if a camchain exists

omni_radtan and equidistant ship with a Kalibr camchain for these cameras, so the calibration is a single stereo run with intrinsics pinned:

camcalib cam-imu -c examples/configs/tumvi_cam_imu/cam_imu_omni.yaml
camcalib cam-imu -c examples/configs/tumvi_cam_imu/cam_imu_equi.yaml

Two steps, otherwise

double_sphere, eucm, and fov have no shipped camchain, so fit the camera first and feed the result in:

camcalib calibrate -c examples/configs/tumvi_cam_imu/camcalib_cam0_ds.yaml
camcalib cam-imu   -c examples/configs/tumvi_cam_imu/cam_imu_ds.yaml

Fit cam0 alone for the wide-FoV models

Stereo camera-only fits diverge on this dataset for double_sphere and eucm — the fx–xi / alpha basin. The single-camera fit is stable, so the recipe fits cam0 only and runs a single-camera camera–IMU solve. See Camera model.

Results — imu1, 45 s window

Model	Mode	Reproj	cam0 lever (mm)	Baseline	td
omni_radtan	stereo, pinned	0.273 px	[45.0, −68.7, −43.2]	100.8 mm	−0.73 ms
equidistant	stereo, pinned	0.277 px	[44.7, −70.1, −43.7]	100.8 mm	−0.72 ms
double_sphere	cam0, pinned	0.275 px	[44.3, −68.5, −43.9]	—	−0.71 ms
eucm	cam0, pinned	0.275 px	[44.3, −68.3, −43.9]	—	−0.71 ms
fov	cam0, pinned	0.288 px	[44.0, −65.2, −44.0]	—	−0.68 ms
Kalibr (omni)	reference	0.11 px	[46.2, −70.9, −43.7]	100.94 mm	~0

What the table says

Every model lands on the same physical rig. Lever ~[44, −68, −44] mm and td ~−0.7 ms across all six. Model choice changes how easily you get a good fit, not the geometry a good fit recovers.

Agreement with Kalibr: omni_radtan and equidistant reproduce the reference lever to ≤2 mm and the stereo baseline to ≤0.15 mm, with the time offset inside 1 ms.

Across all four bags, omni_radtan holds: baseline 100.8–101.1 mm, lever within ~1 mm, and the freed IMU intrinsics (Dw/Da) consistent to under 1%. That cross-sequence spread is the honest error bar — better evidence than any single run’s residual.

CamCalib’s reprojection error is higher than Kalibr’s 0.11 px because the two are not measuring the same thing: different corner subsampling and robust statistics. Compare the parameters, not the residuals.

Pin the intrinsics in the camera–IMU step

Freeing them (free_intrinsics: true) lets the fx–xi (omni, ds) or alpha (eucm) gauge drift along its equivalence class on a single sequence while reprojection stays flat. Extrinsics and td stay correct, but the intrinsics are no longer uniquely pinned. Freeing the IMU intrinsics (imu_intrinsics: dw_da) is stable.

Reproducibility

Item	Value
Configs	examples/configs/tumvi_cam_imu/
Sequence	dataset-calib-imu1.bag, t0 3.0 s, 45 s window
Board	AprilGrid 6×6, 0.088 m, spacing 0.3, t36h11
Reference	Kalibr omni-radtan-1024/, pinhole-equi-1024/
Detection cache	/tmp/camcalib_tumvi_detcache — re-runs repeat only the solve