Strange calibration results

[João Carneiro]

Hello,

I have previously successfully used the toolbox to calibrate a 3
camera and a 5 camera rig but am struggling to get good results with
my current setup of 6 cameras.
The results are far from expected:
- cameras' position between each other seems correct (at least in some
of the calibration attempts) but they are pointing to the wrong
direction
- internal parameters strange including principal point coordinate
outside the image and wrong focal length
- Can achieve a low reprojection error (0.44px) with a low ransac
inlier tolerance

Changes to my configuration files failed to produce better results.

Any help would be appreciated.
Regards

[Andrew Straw]

Hi João,

First, I would make sure you have the global iteration disabled and attempt to identify the non-linear parameters (radial distortion, etc) as a pre-processing step. Also, try re-positioning your cameras slightly (even just a few mm). Doing these things has generally stopped my convergence issues. I think (but am by no means sure) that the underlying algorithm may be subject to the problems specified in:

Oliensis J, Hartley R. Iterative extensions of the Sturm/Triggs algorithm: convergence and nonconvergence. IEEE transactions on pattern analysis and machine intelligence 2007 Dec;29(12):2217-33.

They also specify a solution, which would be great to get implemented in some open source software.

-Andrew

-- 
Andrew D. Straw, Ph.D.
Research Institute of Molecular Pathology (IMP)
Vienna, Austria
http://strawlab.org/