[ptx] automatch: another attempt at semi-automatic panorama stitching

Pablo d'Angelo pablo at mathematik.uni-ulm.de
Wed Jan 7 07:05:00 GMT 2004


On Wed, 07 Jan 2004, Bruno Postle wrote:

some random thoughts below:

> I've played a bit more with my automatch script for finding
> stitching control points.
> 
> -  I tried the optimisation Pablo suggested, it got slower.  I guess
>    I'm not very good at this sort of thing :-p.

Strange. I've always thought perls hashes were very optimized.

> -  I figured that by just using the first 40 feature-points (instead
>    of all 100 points provided by klt) gives a speedup of about 15x.
>    It now runs in about 20 seconds instead of 10 minutes, though it
>    only detects 40% as many points.

You can ask klt to produce only 40 points. maybe thats better, because it
will take the ones with the highest score only. I don't know if klt writes
them to the file sorted by score.

> -  It is incredibly sensitive to accurate lens correction, I tried
>    to match uncorrected photos and got no results at all (note that
>    it can be modified to take account of lens parameters, but they
>    will still need to be known beforehand).

Well, that was to expect if you consider similar points with an accuracy
of +- 1.5 pixel. Distortion is much higher than that. maybe one could think
about introducing a error region, depending on the pixels radial distance
from the center. would probably result in less robust matching, if is not
checked that the shifts used are created by a similar systematic error.

> The process for determining which control-points to use is purely
> statistical, but the thresholds I am using are entirely arbitrary.
> These decisions about which points to discard would benefit greatly
> from a proper statistical understanding.

Hmm, I'm not very good at statistics, so I leave the comments to the experts
;)

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


More information about the ptX mailing list