On 29 Dez., 14:09, Rogier Wolff <[email protected]> wrote: > On Wed, Dec 29, 2010 at 07:42:12AM -0500, Yuval Levy wrote:
> > ... and weighting would help in this and many other cases. For example in > > the > > past the wish has been expressed to discern between user-generated CPs and > > computer-generated CPs. > > Indeed. And a user-preference can set the default weight for both of > them. Some users trust themselves, others trust the computer > better. :-) User input is resource-intensive. If I actually set a CP, I'm rather sure that it belongs there and I trust my vision. I invest time and effort. This should be noted by the program - and honoured by a high weight. > > instead of commenting out CPs, you would set them to a weight of 0. > > Right. Or at least very low. If the feature pops up in hugin, I'll adapt my script ;-) > > Still: I don't think CPs are the ultimate tool for the image > > alignment process... I agree. But what else do we have? I'm asking this because I'm genuinely curious and at a loss here. > > What I think should be possible is that you optimize the cross > correlation of say a 10x10 area of the image with another image. This > is computationally expensive. This would only be feasable to do for > example for a 5 pixel radius of a point 20 pixels north of a > controlpoint. I've tinkered with cross-correlation. It is extremely sensitive to small shifts, so you have to work on interpolated data and nudge them by small increments until you land in the minimum (if you're lucky). Cross-correlation is also very sensitive to (even small) distortions. If you wrestle with it for a while you become so frustrated that you have to use SIFT feature points for a while to take heart again - at least that's my experience. It sounds like a good idea until you get down to actually trying to get it to work. There was an interesting article 'MIKOLAJCZYK AND SCHMID: A PERFORMANCE EVALUATION OF LOCAL DESCRIPTORS' where cross-correlation is discussed together with SIFT and some others: http://lear.inrialpes.fr/pubs/2005/MS05/mikolajczyk_pami05.pdf Feature detectors may still well be the way to go - after all it's a limited understanding of what they do if one just reduces them to the 'control point' one can derive from them. > I can't think of how to get an initial point to start working from > besides asking the user or doing the feature detection thingy.... or use two feature detectors. My idea would be to use the fastest available feature detector, pick it's best few results in a ROI and then verify that these locations are indeed corresponding with a heavy- duty, most-likely-to-succeed detector. Kay -- You received this message because you are subscribed to the Google Groups "Hugin and other free panoramic software" group. A list of frequently asked questions is available at: http://wiki.panotools.org/Hugin_FAQ To post to this group, send email to [email protected] To unsubscribe from this group, send email to [email protected] For more options, visit this group at http://groups.google.com/group/hugin-ptx
