Reprojection errors with April Tags + DLT vs MultiCamSelfCal

[Andrew Straw]

Hi,

(In the interest of keeping the Github issue tracker focused on bugs and conversation here in this google group, I'm transferring Elhanen's question at https://github.com/strawlab/strand-braid/issues/6#issuecomment-886746126 ):

"Should I expect similar reprojection error values as with MCSC? I got around 1.5 for 800x600 cameras and ~4 for the HD cameras." The question refers to usage of the approach here github.com/strawlab/dlt-april-cal versus our traditional github.com/strawlab/multicamselfcal

I have not done side-by-side comparisons of the same rig with calibrations done the two different ways. But indeed I would expect the April Tag calibrations to have a slightly higher reprojection errors than the MCSC approach for a couple reasons:

- The MCSC camera model has more parameters. (Actually I think this was an unintentional oversight on the part of the original authors. It has multiple in-theory redundant but in-reality distinct - and distinctly used - values for things like focal length and optical center.) As a model, it has more degrees of freedom and thus, all else being equal, will fit the same data with less error, particularly if you are not doing cross validation.

- The DLT routine in use is a bare-bones variant which the Hartley and Zisserman textbook calls unstable, but I have never gotten around to improving it.

All that said, it's not clear to me what the effect on tracking performance will be. I certainly advise to be cautious when make comparisons between data from two rigs calibrated with two different ways - as the errors may be different in those cases. Also, it is possible that, if indeed the errors are different, that the tracking performance will also be different.

I hope that's useful.

-Andrew