[ptx] autopano #2, using SIFT

[Sebastian Nowozin]

Hi Pablo, *,


On Fri, Jan 16, 2004 at 06:55:59PM +0100, Pablo d'Angelo wrote:

> On the practical side: so far we have many different languages, and no nice
> codebase. Is .NET really so much better than C++? Is is it hard to port your
> code to C/C++? People complain about the wxWindows dependency of hugin, so I
> don't want to add yet another big dependency if its not really needed.
> Actually all the project/reading writing is already implemented in C++, as
> is an interface for the remapping functions and other stuff, so I would
> prefer if we could use the same codebase.
 
I agree about that it would be unwise to split/port any solid established
codebase, such as Hugin is. My C# implementation is just meant as prototype
again, and it actually was a pleasure to program in C#, compared to C and C++.
I hope Alexandre and me can come up with a C++ implementation and can do lots
of performance tuning along the way. I added a small feature yesterday that
allows for convenient downscaling of an image, and now I can get around 75
precise keypoints with atound 5-10% false positives in two 5 Megapixel images
in 30 seconds using 640x480 images. Using the original image and precise
nearest neighbour searching it would takes ages but yield around 20000 precise
matches with < 1-2% false positives. So, keypoint starvation will never be a
problem.


On the long term, I think it may be better to leave wxWindows and turn to
something like GTK+ or even GTK# for the GUI while using a very portable C++
library and utilities as backend without any non-trivial dependancies. I have
done some work with GTK# now and it is really maturing and such a pleasure to
program with. But this is just a suggestion.


> I'd really like to add feature detection stuff to hugin and a standalone
> program. But I understand if you want to stick to C#.

As long as I have not completely nailed the algorithm I would prefer modding
the C# stuff, but once that is complete, I am all for porting (although it will
be boring, maybe some other people familar with the hugin code can help to
avoid reinventing the wheel).


> Something I'd like to see in any case is a clone of the "keypoints" program,
> available on Dr. Lowe's site :)
 
Thats exactly what my programs do right now. Plus using BBF matching instead of
the vanilla O(N^2) euclid. And some small extra filtering on the match results
(removing repeating features).


> > 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.
 
> Could you ask for a reference to the patent? I still haven't seen it...
 
He has not answered my mails yet (and I would understand if he does not, I
would not want to bother with such legal fuzz either).


> ciao
>   Pablo

ciao,
Sebastian

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