[ptx] Even more SIFT successes

[Sebastian Nowozin]

Hi *,


You all must be becoming pretty tired of hearing of yet another success ;-)

I hacked up a fast multimatching tool to my SIFT implementation, so the
matching complexity for n pictures is down to O(n). In practice its around half
a minute for 6 pictures, 5MP each with 1000 features. All keypoints are put
into a large k-d tree which is used to create a matching list. The list is
partitioned into image pair match lists, which are filtered on keypoint match
count (>= 3 && 12 best selected).

Also, I wrote a small .pto output program and this is a result image without
any modifications on the control points (infact, they are all 100% correct).
This is amazing! :)

	http://user.cs.tu-berlin.de/~nowozin/sift/100pct_AUTO.JPG  [2.9Mb]

I ran the optimizer afterwards, which changes the average control point
distance from 0.9 to 0.8, with the worst being 1.9. (This is as good as my best
manual stitches)


Btw, Pablo, I found a nice criterion for the suitability of a match:

 w(dist1, dist2) = dist1^e1 * (1 / (dist2 - dist1))^e2

Where dist1 is the distance between the target keypoint and the first match,
and dist2 is the distance between the second best match and the target
keypoint. e1 = e2 = 1.0 work nice for me, but one could tweak the ratio of "how
close" against "how closely followed" by adjusting them.


ciao,
Sebastian

PS: I will try to give out a usable version as soon as possible, but there is
    still that patent issue with SIFT...

-- 
nowozin at cs.tu-berlin.de --- http://user.cs.tu-berlin.de/~nowozin/