Calibration issue

finn

Hey,

we have about 200 oak-d pro PoE in use. We use the factory calibration and every then and now stumble across bad calibrated cameras (see images below).
Looking at the extrinsics from CAM C to CAM B we see a suspicious translation vector .

R = [0.99997956, 0.0055608875, -0.0031641622;
-0.0055211708, 0.99990761, 0.012423742;
0.0032329522, -0.012406022, 0.99991781]

t = [-7.2643051; 1.68738; -1.025482]

Flashing new extrinsics from another camera closer to

t = [-7.5; 0; 0]

yields better but not perfect results.

Is it advised to perform a calibration ourselves?

Would dynamic calibration be useful in our setting in a rather texture poor environment?

Oskar-Luxonis

Hey finn - The translation vector itself looks reasonable: its magnitude is 7.53 cm, which is very close to the nominal 7.5 cm baseline. I would avoid flashing calibration data from another unit, as its intrinsics and distortion parameters are unit-specific.

For units that remain miscalibrated after restoring the factory calibration, you can also perform a manual calibration. The process is straightforward and documented here: Manual Calibration.

Dynamic Calibration can also help compensate for a changed physical alignment, but it needs a feature-rich scene across the FOV. In a texture-poor environment it may not collect enough reliable data, so we would recommend running it during commissioning while viewing a textured scene, then validating the result before deployment.

Thanks,
Oskar

finn

Hey @Oskar-Luxonis

yes your right

|t|≈7.5cm

but y and z component of more than 1 cm seem very high and are not observable on correctly calibrated cameras.

Could this be used to filter for miscalibrated cameras? Or is there another way? I want to avoid recalibrated all devices.
Furthermore, I do not have physical access to the camera in the image above and therefor don't have a chance to perform manual calibration. Can dynamic calibration help in this case?

Thanks, Finn

Oskar-Luxonis

finn - I agree that the Y/Z components are unusually large compared with your good units. You can use those values as a first screening filter within the same OAK-D-PRO-POE revision, but I would not use a fixed threshold as the only pass/fail criterion.

A more reliable check is a calibration-health test based on reprojection error. In DepthAI v3, Dynamic Calibration can report calibration metrics and confidence, but it needs live stereo images with sufficient texture. It can therefore help remotely only if you can run it on the device and collect data from a feature-rich scene.

Could you confirm your DepthAI version?

Thanks,
Oskar

finn

Hey @Oskar-Luxonis
We have been using depthai v3.7, but will now upgrade to 3.9.

Further investigation of the issue lead me to a camera with a worse calibration. Looking at the factory calibration I get:
{x: -7.5871477127075195, y: -0.056533243507146835, z: -0.058885782957077026}
for the user calibration I get:
{x: -4.399874687194824, y: 0.040574610233306885, z: -6.1814985275268555}
restoring the factory calibration leads to much better results.
We have not been changing calibration. I therefor suspect dynamic calibration, which, if I understand correctly, is turned on by default. Could this be true?

Oskar-Luxonis

Hey finn - yes, this is possible. DepthAI v3.7 enables AutoCalibration in ON_START mode by default. For supported stereo streams, it can save a validated result as user calibration. It does not change the factory calibration.

We cannot yet confirm that AutoCalibration caused this specific user calibration. But the difference between the factory and user calibration shows that a user calibration was written to the device.

As a test, please restore the factory calibration and explicitly disable AutoCalibration in your application with pipeline.setAutoCalibrationMode(dai.Pipeline.AutoCalibrationMode.OFF). Then check if the user calibration stays unchanged.

Since your normal scene has low texture, I recommend that you keep AutoCalibration disabled in production for now. You can use manual calibration for the cameras that still show poor depth after the factory reset.

Thanks,
Oskar