[ptx] autopanos?

[Sebastian Nowozin]

Hi :)


On Thu, May 06, 2004 at 08:09:48AM +0200, alexandre jenny wrote:
 
> Autopano Alexandre Jenny (That's me ;-))
>  + Speed, it's c++.
>  + Very low memory usage.
>  + Ransac check (removes many false point).
>  - Accuracy. Around 5 % of false points still remains.
>  - Not open source yet.
 
Yes, RANSAC is a great plus, I am looking into implementing some further
postprocessing, too. It really helps.


> Autopano_sift Sebastian
>  + Accuracy. It seems to be more accurate.
>  + Open source.
>  - Speed.
>  - Memory usage.

I do not know where the additional accuracy comes from, but I improved a lot of
the SIFT code in the latest version. Especially for diffuse surroundings and
soft gradients, the filtering is stricter, which leads to fewer keypoints and
better reproducability. Also, I think the very simple measure I implemented
about the quality of a single match really helps, as it weights of the overall
quality of the match (the absolute distance between the two keypoints), and the
relative stick-out-from-rest quality (the ratio of distance to the next second
best match). As some small helper, the matching also removes join-matches,
which appear in repeating image elements, where many keypoints from one image
are mapped to the same in another.

The speed and memory usage is really a pitty, yes. I like to think its not a C#
thing, although the floating point code generated by the Mono IA32-JIT is
really not the best. The only part where the speed is really bad is at the
matching. In the latest version I replaced a lot of FPU intense code with
integer code, which doubles the speed. For the memory, I plan to change the
order of the octave processing to a linear one, which would reduce the peak
memory usage. But still, images eat memory :) Maybe replacing all double's with
float or some fixed-point 16bit format will help. Mhhh.. if we would just have
more time for all this optimization stuff ;)


> Am I right ?
 
Yes. :)


> Bye
>   Alexandre 
 
> We should use some test case to make a real comparison ... Hum ...
 
Not just for comparison of our programs, but also for relative versions of the
same program, changed threshhold values and - more importantly - so we can
discuss results on this list.

bye,
Sebastian

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