[ptx] autopano #2, using SIFT

[Pablo d'Angelo]

Hi!

I have just finished the first version of "panosifter", that is autopano on
steroids.

Give it a bunch of image files, and it will create multiple hugin projects
from them. All this is based on the SIFT feature detector research demo by
Lowe, which is available for research purposes only, so we cant really use
this like it is.

But I hope that we can replace it with our own feature detector soon :)

Basically it runs the O(n^2*m^2) matching step with m small images, that
contain around n=50-200 features and builds a graph of all images, which is
searched for subgraphs (separate panoramas)

Only the overlaping images are subject a more expensive matching. This works
quite well. The matching algorithm is still quite simple, but the SIFT
features provide a very solid base. I haven't run into problems with
misdetections yet, even scences with many similar structures, like old
churches are handles quite well. The occasional outlier doesn' really hurt
if there are correct points.

I have tried to make the code modular and easy to understand, but I have
hacked on today and yesterday, so maybe there are some things that could be
made better.

But it should be easily possible to integrate other implementations. If a
external keypoint program with the syntax as the one from lowe is availabe,
no further integration is nessecary :)

Btw. Sebastian, do you want a CVS account?

here's the usage:
-------------------------

$ panosifter -h
panosifter: create one or more .pto files, from multiple images.

   Does a scouting feature detection on all image combination, with
   very small images, only does the more expensive matching
   for the confirmed links.
   All the hard work is done by David Lowe's SIFT feature detector

Usage: [options] -o pto_prefix image1 image2 image3 ...

  [options] can be:
     -o prefix   # prefix of the created hugin project files
     -n number   # number of features per overlap, default: 20
     -v number   # HFOV of images, in degrees, Used for images that do
                   not provide this information in the EXIF header
     -r number   # downscale factor for scouting. default: 0.05
                   increase if images are wrongly categorized.
     -s scale    # downscale images before final sift matching.
                   default: 0.25
                   while this seems crude, it is needed for speed
                   and still gives good results


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