[ptx] libsift control points

[Ed Halley]

On Fri, 2004-10-08 at 17:07, Eric Engle wrote:
> So I backed up and took a closer look at the control points.  I did a
> pixel-level comparison between features present in both images and found
> that the control points were NEAR the same features but were between 5
> and 15 pixels translated.  Unfortunately, the degree and direction of
> translation is different at each control point.
>
> The images are snapshots of the Martian surface.  They are at slightly
> different contrast levels since the satellite took them during different
> orbits.  The field of view is extremely narrow, 3-4 degrees.

"What we've got here is a target-rich environment."

I can only speculate here, not having seen the images, but I think the
algorithm is finding a TON of potential sift-able features which all
look basically the same, and then just not picking the correct set when
they're being correlated.  The correlation process tries to find the
most obvious features found, and then select other features that have
approximately the same relative angular layout.  It then discards the
rest of the features, which hides their potential value in matching up
more accurately.  It does not try all possible combinations exhaustively
by brute force.

The same would happen if you took two pictures of shag carpet or
sandpaper, the algorithm would pick two similarly dentrite-shaped sets
of features, without being able to accurately judge if they're the SAME
set or not.

-- 
[ e d @ h a l l e y . c c ]