Hallo, Am Dienstag, 24. Juli 2018 22:13:04 UTC+2 schrieb Erik Krause: > > Am 24.07.2018 um 21:01 schrieb [email protected] <javascript:>: > > > I found an old thread on the panotools wiki. > > > > https://wiki.panotools.org/User:Klaus/Improving_Hugin > > There probably is one problem we didn't consider then: Higher order > polynomials tend to overfit.
Yes, this is a problem in principle. But if one uses enough CPs for a given number of images, a good part more than the degrees of freedom / fitting parameters, then the fit is overconstrained and overfitting is not an issue. Of course make sure the CPs are reasonably spread over the image areas. That might be the reason why professor > Dersch choose the current model (he's a mathematician after all). Side informations: as a physicist I've also been rigorously trained in mathematics. Professorship is no vaccine against getting something wrong ;-) With due respect for Helmut Dersch and his pioneering work in panoramics, devising and coding some sterling software along the way, I think he just overlooked the fact that his parametrisation has function terms that are non-holomorphic at vector-r = zero (maybe some slight prejudice towards Fachhochschule there ...) Or maybe he did not care; a and c do give some small improvement, and being out of alignment by 3 or 5 pixels in the image corners is no big deal if your image seam runs elsewhere. It is interesting, that NASA people (do not find the reference right now, either for star or for rocket tracking) have also calibrated their lenses with multiple images. They use only odd polynomials. But in the paper they do not give a reason for that. Maybe it seemed obvious to them.... > At least that's the opinion of Joost, the maker of PTGui (where this > discussion pops up from time to time, too). > > There's nothing gained if we have a "correct" lens distortion model that > can't be safely optimized... > I do think that two more parameters, for r^5 and r^7 (if one elects to keep a and c switched on as fit parameters) do not create optimisation instability (of course start optimisation without barrel distortion) when there are typically a few dozen optimisation parameters. Of course the proof lies in the puddin. Let's give it a try. Best regards Klaus P.S. What amazed me, years ago, take a rectangular lens model r~tan phi, add r^3 and r^5 and r^7 terms, and you can get a near-perfect description of a fish-eye lens with its r~phi property. -- A list of frequently asked questions is available at: http://wiki.panotools.org/Hugin_FAQ --- You received this message because you are subscribed to the Google Groups "hugin and other free panoramic software" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. To view this discussion on the web visit https://groups.google.com/d/msgid/hugin-ptx/61c0ace1-377e-4ffc-b8a6-75e8fd5ccafe%40googlegroups.com. For more options, visit https://groups.google.com/d/optout.
