Round-trip validation

Reprojection error tells you a calibration fits its own observations. It does not tell you the calibration is right. Round-trip does: apply the calibration to the data it came from, then calibrate again. If it was correct, there is nothing left to find.

camcalib roundtrip -c cam_imu.yaml

One command runs calibrate → apply-calib → recalibrate → report. It exits non-zero on FAIL, so it belongs in CI.

What PASS means

Two conditions, and they pull in opposite directions — which is what makes the test informative:

Collapse. On the corrected bag, every IMU’s Dw and Da must go to identity and b_g / b_a to zero. The first pass claimed to have measured those; if it was right, correction consumed them and there is nothing left.

Preserve. The camera and IMU extrinsics — and the stereo baseline — must not move. Correction changes timing and inertial scaling, not geometry. A baseline that shifts means the first solve traded geometry against something else.

Residual td is reported as a warning, not a failure: how much of a time offset survives correction is observability-dependent.

Flags

Flag	Effect
-c, --config	The cam-IMU config to validate
--out	Report path
--bag	Override the config’s bag
--keep-corrected	Keep the intermediate corrected bag
--keep-bias	Passed through to apply-calib

The corrected bag is given a distinct name — so it gets a fresh detection cache rather than silently reusing the original’s — and is deleted afterwards unless you keep it.

Confirmed results

Rig	Collapse	Preserve
virig 3cams	all four IMUs’ Da → identity	baseline 69.86 mm stable
Monado / MIC05	Da 1.06 → 1.00, b_a −0.25 → −0.01	baseline 133.8 mm stable

When it fails

A FAIL is a real finding, not a flaky test. The usual causes:

• Wrong camera model — the intrinsics absorbed error that reappears once timing is corrected. Try another distortion model.
• Wrong IMU convention — see IMU model.
• An interpolated accelerometer corrupting td. This is the T265 case; gyro_only on that IMU, or estimating t_ga, is the fix.
• Insufficient excitation — the parameter was never observed, so correcting by it does nothing consistent.

Same entry point as studies/roundtrip.py.