[ptx] autopano: first attempt at semi-automatic panorama stitching

[Bruno Postle]

On Fri 02-Jan-2004 at 04:32:19PM +0100, Pablo d'Angelo wrote:
> On Fri, 02 Jan 2004, Bruno Postle wrote:
> 
> > However, if you have a photo that contains 'features' and you know
> > roughly the type of lens and fov, there is an attribute that is
> > independent of roll, pitch and yaw that can be calculated precisely:
> > The angular distance between any two features in a single image.
> >
> > Say you have an image with three features; there will be three of
> > these angles that can be calculated - If another photo has three
> > features with the same three angles, then the two photos are quite
> > likely to actually correspond at those points.

> Using just 3 features will be very unreliable, because a single feature
> might not be detected the same way in the other picture, or it's not on the
> image at all. 

If klt can identify 100 or so likely features in any one photo, we
only need a handful of them to be common between any pair of
pictures.

Matching might proceed like this:

Say we have a list of 100 features per image, there are only 4950 or
so angular distances to calculate.

Most of these angular-distances can be instantly discarded because
the value will be too close to the angular resolution of the camera
to be useful, or because the value is unique between the set of
photos you want to match.

Create a table of distances for each photo, each item in the table
is (angular_distance, node_a, node_b).

For each of the pair of photos you are trying to match, try and
create a network of nodes that has an equivalent geometry:

-  The simplest network is two nodes and one distance-link, there
   may be a lot of these with a good match for each pair of photos,
   they will be very easy to find.

-  Most of these single-link network matches will be bogus, but they
   can be tested by trying to expand the network to three nodes and
   three links, the search space for this is very small since there
   are only 98 remaining nodes to check.

-  This 3-node 3-link network can be tested by trying to expand it
   into a 4-node 6-link network etc..

All this assumes that klt is capable of matching at least a small
proportion of features consistently between two photos.

-- 
Bruno