Monado

The Monado calibration sequences (TUM MIC) come from a Valve Index headset rather than a research rig, and ship with Monado’s own published calibration. That makes them a second opinion on TUM VI from a different toolchain — and the dataset where t_ga and round-trip validation earn their keep.

The rig

Sensor	Detail
2 × monochrome cameras	Stereo fisheye
IMU	`/imu0`
Reference	Monado / Valve Index, baseline 134.0 mm

Board: AprilGrid 6×6, tag_size 0.03 m, tag_spacing 0.3, family t36h11.

Getting the data

EuRoC mav0 layout — no conversion needed:

tum_mic/mic_calib/MIC05_imucalib2/mav0/

MIC05 is the canonical sequence; MIC01, 02, 03, 07, 08, and 11 are also configured, which is what makes a cross-sequence spread available.

Configuration

input:
  format: euroc
  path: /data/tum_mic/mic_calib/MIC05_imucalib2/mav0
  imu_topics: ["/imu0"]
  window_mode: motion        # auto-pick the most-exciting 80 s
  segment_length: 80.0
  cache_dir: /tmp/camcalib_detcache

board: {board_type: apriltag, rows: 6, cols: 6, tag_size: 0.03, tag_spacing: 0.3, tag_family: t36h11}

camera_input:
  path: examples/regression/tum_vi_mic/MIC_camchain_v2.yaml

solve:
  max_iterations: 40
  init_kwargs: {keyframe_hz: 2.0, scale_rel_tol: 0.3}
  estimate_t_ga: true
  sigma_t_ga: 0.010

estimate:
  camera: {intrinsics: false, tr: false, td: true, extrinsics: true}
  imu:    {intrinsics: dw_da, model: calyx}

output:
  path: results/regression/tum_vi_mic/mic05/cam_imu.json

reference:
  label: Monado Valve Index
  baseline_mm: 134.0

Two choices worth noting: window_mode: motion picks the most-excited 80 s automatically rather than making you hunt for it, and intrinsics are pinned to the committed camchain so the solve concentrates on extrinsics, td, and the IMU.

Running it

camcalib cam-imu -c examples/regression/tum_vi_mic/cam_imu_mic05_tga.yaml

Validation — round-trip

The result is not accepted on its residual. It is applied to its own data and recalibrated:

camcalib roundtrip -c examples/regression/tum_vi_mic/cam_imu_mic05_tga.yaml \
  --bag MIC05.bag

Quantity	Before	After correction	Verdict
`Da`	1.06	1.00	collapsed ✅
`b_a`	−0.25	−0.01	collapsed ✅
Stereo baseline	133.8 mm	133.8 mm	preserved ✅

Both conditions hold at once, and they pull in opposite directions — the accelerometer scale and bias the first pass claimed are consumed by the correction, while the geometry does not move. Against Monado’s published 134.0 mm, the recovered 133.8 mm is 0.2 mm out.

See Round-trip validation.

`t_ga` on and off

Both variants are configured — cam_imu_mic05_tga.yaml and cam_imu_mic05_notga.yaml — precisely so the effect of estimating the gyro↔accel offset can be isolated on real data rather than argued about. See IMU model.

Reproducibility

Item	Value
Configs	`examples/regression/tum_vi_mic/`
Sequence	`MIC05_imucalib2`, motion window 80 s
Board	AprilGrid 6×6, 0.03 m, spacing 0.3, `t36h11`
Camera intrinsics	pinned to `MIC_camchain_v2.yaml`
IMU model	`calyx`, `dw_da`, `t_ga` estimated
Reference	`monado_reference_calibration.yaml`, baseline 134.0 mm