experiments with Dr.Lowe's SIFT demo Re: [ptx] autopano #2, using SIFT

[Pablo d'Angelo]

On Fri, 16 Jan 2004, Alexandre Jenny wrote:

> > Something I'd like to see in any case is a clone of the 
> > "keypoints" program, available on Dr. Lowe's site :)
> 
> Keypoints just gives keypoints : it does not the matching. And I thing this
> part is
> Much more interesting than the features.

I have used the demo program from Lowe, and it gives fantastic results.
I haven't used a fast nearest neighbour search, nor the hough transform,
just the simple euclidian distance measure on the feature descriptor, as
used in the example match.c program.

I've added an interface for keypoints into hugin, one can select some images
and all of them are matched with sift features. (with 4 times lower
resolution), the results are very good and despite my simple approach of
matching everything with everything, the matching speed is acceptable,
taking ~10 seconds for every pair + the time for calculating the sift
descriptor. It also gives plenty of good matches. typically 50 - 700 per
pair. First I didn't have a threshold, so I ended up with several thousand
control points for a 5 image pano. ;)

In short: the SIFT features work really well, even without sophisticated
algorithms on top of them. Hope we can integrate a similar feature detector.

Hmm, PTOptimizer is quite slow with 2000 control points.

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