Hacking the lens calibration process?

I do love the camera. The lens calibration wizard is “ok”, but not really working so well TBH. I have fisheye lens mode and got decent scores, but the result is still off by several mm is central area and wildly off- like 30mm- near the edge. It’s not the reproducibility of the lid position, I ran all this with the lid interlock disabled so I left it open and never touched it. And clearly not an offset prob as, even after cal, the edge of the field is bowed in by that ~30mm along the edges with uncompensated barrel distortion. Also most of the error was in left/right offset which would not be the axis with the most error if the lid were not open in the same way.

Instead of engaging LB devs to try something that might be better, it occurs to me an independent coder can do this. Like make a document with a arbitrary grid of 24 crosshairs on a black layer to burn, then take a shot with the camera, and manually create red crosshairs on the same document where the burns appear on the camera vs the black crosshairs in actual coordinates. From there, a Python script could take that .lbrn2 file of camera points, and the machine’s lbdev file, and recalculate both the lens calibration and alignment parameters and write new lens parameters back to the .lbdev file.

That would replace both the camera calibration and alignment steps, which might be able to yield better accuracy. Is that possible? Can we get a description of how the lens parameters in the XML work?

Or is the problem in barrel distortion near the edges being tens of mm off a limitation of the algorithm that tweaking the available compensation fields with better numbers wouldn’t necessarily make better?

Without having the necessary programming knowledge, I would still think that the perceived inaccuracy is not related to the calibration. Otherwise the same problem will occur for all users.
I do not want to defend the calibration program itself, I have sometimes spent too much time with it to get the desired result, but on the other hand, once I had good numbers and was able to complete the entire process, the photo-optical system was extremely accurate at me. On my small 200x300 mm (K40) machine, the precision was significantly below 1mm over most of the area. My camera was a noname low resolution type.
On my new machine I do not have a camera installed yet, but am betting on buying a LightBurn “Original” this time. The problem is my OS (Linux Mint) which is not so happy with LightBurn with camera or vice versa. But this too will be resolved soon.

I have the “official” 8MP 4w 120 deg camera and the arducam 16MP which I like better. The lens cal scores well but was quite bowed by barrel distortion in the end. Which, oddly, was in the opposite direction of the original uncorrected feed which bowed outward

So it still overcorrected, by quite a lot! Mostly on the edges, which maybe isn’t used that often but it’s definitely a limitation

This topic was automatically closed 30 days after the last reply. New replies are no longer allowed.