[ptx] autopano #2, using SIFT

[Sebastian Nowozin]

Hello panorama enthusiasts :)


I am about to complete my own implementation of the SIFT image feature
algorithm by Mr. Lowe (http://www.cs.ubc.ca/~lowe/papers/ijcv03-abs.html),
which is the state of the art in invariant image feature extraction,
providing full scale- and rotation invariance and limited variance to affine
transformations and color/brightness changes.

Enough said, here are the screenshots of only very primitively (ie
non-geometric, just join/split and 1st/2nd-matching) filtered correlations:
http://user.cs.tu-berlin.de/~nowozin/sift/

While one small piece of the algorithm is still lacking, most other things
including the BBF search of the paper is implemented. The next step is the
Hough transform matching and a way to digest and output hugin stitcher files.
I hope to work together with Alexandre on this parts :-)


I plan to release it under the GPL soon, but want to discuss this with Mr. Lowe
first, as you all know the SIFT algorithm is patented. The implementation is
100% C# btw. A 640x480 image takes around 30s processing for the feature
extraction and it takes roughly 40s to match two such images. But there is lots
of room for improvement left. The most important problem to get this into
practical use is the RAM demand, which is around 400Mb for a 640x480 image
feature extraction.


ciao,
Sebastian

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