I gave up trying to calibrate the lens while attached to my laser, so I put the test pattern on a 4’ x 4’ sheet of flat white melamine/mdf board, sitting vertically against a wall with my camera on a tripod. I have tried using the flat brown side as well. I have an 8MP ELP camera module that has not worked; and is supposedly the same camera that Lightburn are selling as 8MP N 75°, and a new ELP 16mp 100° FOV Aspherical Manual Focus, which has ALSO failed to calibrate during this testing. The camera I am currently testing is an AverMedia Cam340 - 4k, 85°H/55°V FOV Aspherical lens (This means distortion free - NOT a fisheye) The only reason I’m testing this camera so extensively, is because the 8mp and 16mp cameras I tried have had even lower capture recognition rates.
This is the 3rd time through my testing to get a solid, repeatable lens calibration with a specific camera. I have been testing this for almost 7 hours today.
I got so frustrated I wrote myself a program to track how many times I’ve clicked the ‘Capture Image’ button in Lightburn trying to get my camera lens calibrated. Right now the counter is on 2843 clicks… with ZERO pattern recognitions beyond the first image.
I have the correct test pattern with the round dots offset - 4 rows of 6 dots alternating with 4 rows of 5 dots. I have tested the camera from 15cm to 1 meter from the target, even putting tape lines on my monitor screen over the top of the live camera view at one point to guarantee the lens calibration target properly fills the center of the 3x3 grid like others have said is required, which places the camera about 38cm from the target. I even have external fill lighting set up to provide even lighting so the camera wont see any washed out hotspots or shadows on the target.
In my second round of testing, before my current setup, I tried changing orientation and fisheye settings, and I got some odd results during roughly 1000 captures, give or take a couple hundred. I have sequential screenshots of Lightburn giving me a recognition score of 0.68, then 8888.0 (with no image), then a clear capture image with No Pattern Recognized, a 9.61 with a badly fisheye distorted image, and then a recognition score of 85047208.00 with a solid gray capture image. Really? Should this score even be possible when you want something that is less than 0.9?
Maybe a half a dozen times, the ‘recognized’ image was INVERTED from the live camera view, both when the paper target was physically inverted on my board (the capture image was right side up) and when the target was right side up, the capture image was then inverted and gave me a score in the 1.xx range.
I’ve set this same camera to fisheye, and I’ll get small bursts of recognitions with a capture image, and those captured images sometimes look perfect like a flat photo that should be perfect (with abysmal recognition scores), and sometimes look like they came from a 360° camera with a score in the 1’s and 2’s.
Whatever capture recognition processing is being done, also does this random fisheye warping to a much lesser extent in the standard lens setting. Probably 5% of the captured image located around the borders is curved sometimes but not always. it’s like every attempt you click “Capture Image”, it uses a different variation of image processing and recognition, and it swings wildly in what it changes instead of using incremental shifts.
In the VERY few instances where I got an acceptable score below 0.9 for the center capture, I spent DOZENS of attempts to get the second capture position and had zero recognitions, so I started the process over.
Camera lens calibration should NOT be this unreliable. camera lens FOV are not infinitely variable here. You should be able to tell Lightburn that you have a camera with a resolution of X, and an FOV of Y, and it should be able to tell you “set your camera to approximately this height”, and then adjust what type of recognition algorithm it’s using so it isnt making attempts using garbage where every image capture is distorted in new LSD inspired ways so that lightburn can’t find the dots that even 20 year old OCR software can find.