Why does the Stereo Camera Calibrator App provide different results on different MATLAB versions?

Hello all,

For the images given here, I used the Stereo Camera Calibrator App (see the attached calibration session).

On R2021b, the app provided a calibration output (parameters). Still, using the exact set of calibration images on the app running on R2023b, the app was not able to solve for calibration parameters and through the following error:

Does anyone know why it is happening?

Thank you,

Elad.

Answer 1

[Matt J]

LIkely because the patterns are too similar across images, like the error message says. This would lead to ill-conditioning of the calibration problem, making any solutions you got in the past unstable..

[Elad]

Thank you Matt for the quick response.

You may not have seen that I wrote that the calibration worked in R2021b. I have even attached the successful session, which can be opened in every version of Matlab.

Nonetheless, it didn't work on R2023b, and I wonder why because this is the version installed on my PC. It didn't work even when I dropped some similar image pairs, so my question remains.

Anyone who wants to see for himself can open the session in R2023b or above and click "calculate".

[Matt J]

You may not have seen that I wrote that the calibration worked in R2021b

No, I did see that. My answer is the same. If you had bad data in R2021b, you may have gotten a solution, but not a stable one (i.e., you had overfitted calibration parameters). Apparently R2023b now includes some extra pre-checks on the quality of the data.

[Elad]

Oh, I see. Cool.

How can I determine if the solution is overfitted? Upon inspection, it appears to replicate the checkerboard quite accurately.

I attempted to eliminate some visually similar pairs in R2023b, but it still didn't work. Have you reviewed the images?

Thanks.

[Matt J]

How can I determine if the solution is overfitted?

Well, it appears the app tries to do that for you in r2023b ... hence the error message. I don't know what image similarity criteria it uses. You could try retaking the calibration images with the chequerboard more tilted and off-centered. The two cameras should be given as distinct perspectives of the chequerboard as you can manage.

Ultimately, the proof that the calibration worked is how well the stereocamera works, once calibrated, on things other than the chequerboard.