[ptx] autopano things

[Pablo d'Angelo]

Hallo Sebastian,

Sebastian Nowozin schrieb am Donnerstag, den 01. Januar 2004:

> I am very happy the mention of my naive "autopano" program a week ago caused so
> much fruitful discussion here and resulted in a far more advanced autopano
> program by Pablo. Although I already consider my original program to be
> obsolete in any way, it at least had some benefit in causing a bit of
> discussion :-)

Well, I just saw that lucas kanade tracker implementation is quite good and
hacked together a program that combined it with the pase correlation. For a
real autopano, a different approach is needed. especially the phase
correlation needs to be replaced by something else, that is aware of the
projections were are using.

the current autopano program is just a quick hack, that works for single row
panos shoot with a roughly constant overlap area. So it is also kind
of a dead end, in the sense that it does only support a small set of use cases
of the pano tools.

> But this is as good as my first approach can become and I won't go this
> dead-end path any further :) Maybe the time to read and understand more
> advanced papers really repays. I hope I can contribute some code once I
> understand some of the SIFT, KLT and other methods used by you guys.

Well, I wouldn't say that I understand everything in that area ;)

I really think, that using the ideas from Brown's panorama paper is a good
start. If its just the SIFT features that are the problem, we should try to
use another feature representation, with similar properties.

Alexandre, could you provide the non-SIFT parts of the source you have
written up to date? Doesn't matter if its matlab only, or the c++ version is
buggy.

> Another small idea I had today which might work independant of the
> feature-matching method used: To get rid of recurring elements (such as rows of
> windows, etc.) one could match the image with itself and record all the
> non-trivial matches. If their match-distance is below some threshhold, both
> features should be removed from the matching.

Hmm, maybe that helps with your approach. On the other hand, one could check
for similar features while extracting the features and remove similar
features. But that might be a bit crude. probably better to let the higher
level algorithms like RANSAC deal with that problem.

The nice thing in autopano is that the phase correlation doesn't need
features, so and the klt searches only a local neighbourhood, so that
repeating structures and stuff like that are handled well, as my tests have
shown.

PS Stan Birchfield has allowed us to use his KLT implementation, I'll
   add the code to CVS soon.

I wish everyone a happy new year!

ciao
  Pablo
--
http://wurm.wohnheim.uni-ulm.de/~redman/
Please use PGP