[Elphel-support] camera image undistortion rectification

[support-list]

Hello Gergely,

We did not try to use others calibration software - we developed our own on top of the aberration measurement/correction program that we made earlier. You can find the details on our blog. We use custom pattern - curved checkerboard as it has more uniform spatial spectrum than the plain one. Registration is based on mixed correlation/phase correlation over 128x128 (or 64x64) pixel area for each grid node (includes multiple of the pattern cells) and results in ~ 0.05 of a pixel error.
The overall calibration is done with rotating/tilting the camera, repeated from 3 positions. The target itself is calibrated - each grid node has a triplet of x,y,z correction from the ideal uniform grid. When calibration Eyesis4pi (it has 26 sensors) software processes about 2-3 thousand images with total of 3-4 millions grid points - that allows to achieve reprojection error RMS of 0.06-0.08 pixels, with worst error of 0.3pixels.

We used 3 different pattern/materials (not to count original paper ones), the current pattern is 3x7 meters and is made of the self-adhesiive  vinyl film attached to the wall - that makes it rigid and the thermal expansion is lower than of the free-hanging plastic panels.


All the software we use is licensed under GNU GPLv3, so it is possible to reproduce the setup. But it may technically be difficult because it relies on a high quality pattern and additional hardware that is hard-coded in the program, so when using different devices some code modification will be needed.

This additional hardware includes:

1 - rotating machine - it is designed primarily for the large Eyesis camera, so smaller cameras can use regular PTZ camera mount (high precision is not required as the software uses mechanical coordinates only as a hint.

2 - a set of 4 laser pointers, controlled by the software - the pointers show 4 different points on the pattern - that allow detection of absolute orientation, during the first fitting pass only images with >=2 pointers are used, so symmetry of the pattern does not cause ambiguity, later images with 1 pointer can be used (when approximate orientation of the camera as a whole (and camera modules in the camera) are known. And at the very end images with no pointers can be used - software tries to match the pattern around the expected one (pattern can be visible by several sub-cameras and some do not see pointers).

You are welcome to come visit us here at Elphel - we can help you calibrate your camera in our new calibration facility.

Andrey


---- On Wed, 10 Jul 2013 09:49:39 -0700 Gergely Debreczeni  wrote ---- 

>Dear Andrey and all, 
> 
> Sorry for returning back to your question so slowly, I was distracted with other issues. 
> Unfortunately we still have the problem of the correctly calibrating the stereo camera. 
> 
>We have a MNC354-2 Stereo Camera with 353-70-50 f=5 mm C mount lenses. 
> 
>What is the best toolkit / distortion model to calibrate these lenses ? 
>We have tried OpenCV stereo calibration utility, and now trying "Camera Calibration Toolbox for Matlab" 
>but none of them gives very good results. (By eye it looks ok, but disparity fails...) 
> 
>For the calibration we used 25 stereo image of a big chessboard from different viewpoints. 
> 
>Questions: 
>- Do you have specific suggestions how to track down the problem ? 
>- Is there somebody having similar lenses and managed to calibrate them successfully ? 
>- What typical errors one gets (order of magnitude of calibration parameter errors) ? 
> 
>thanks a lot in advance for any help, 
>Gergely 
> 
> 
> 
> 
>________________________________________ 
>De : support-list 
>Envoyé : jeudi 6 juin 2013 01:14 
>À : Gergely Debreczeni; Elphel List 
>Objet : Re: [Elphel-support] camera image undistortion rectification 
> 
>Gergely, 
> 
>That camera is not calibrated - it is just hardware that allows to capture synchronized pairs of images, so there is some shift, rotation and lens distortion. The easiest thing you can do to start - capture a pair of images that include fine details at large distance and then manually match the images, shift and rotation is the first thing to process. 
> 
>What type of lenses did you get? 
> 
>Andrey 
> 
> 
> 
> 
>----- 
>No virus found in this message. 
>Checked by AVG - www.avg.com 
>Version: 10.0.1432 / Virus Database: 3184/5888 - Release Date: 06/05/13 
> 
>_______________________________________________ 
>Support-list mailing list 
>Support-list at support.elphel.com 
>http://support.elphel.com/mailman/listinfo/support-list_support.elphel.com 
>