[ptx] Crash on optimise

[Pablo d'Angelo]

On Tue, 26 Oct 2004, Rik Littlefield wrote:

> It may be that there is some subtle error in memory management occuring 
> elsewhere in hugin, that is getting picked up by more strict checking 
> triggered by the MSVC build.  If so, it could be the devil's own time to 
> find without a good memory checking tool.

Yes, should do this sometime. 

> Switching subjects just slightly...
> 
> Have you gotten any feedback regarding whether other people find 
> pairwise optimization to be particularly helpful? 

No, nobody has replied to me yet.

> I confess that from skimming the code I'm not quite sure what pairwise 
> optimization does.  I gather that it starts from an anchor image and 
> works its way outward -- first optimize each image that touches the 
> anchor directly, then each image that touches one of those, etc.  Is 
> that right?

Basically yes.

The algorithm goes like this:

It creates a graph with the images, the nodes are the images and the links
are control points that connect the images.

Then I start from an anchor image, and to a breadth first traversal of the
graph. Everytime I discover a new node (image), it is added to the set of
images to be optimized. The optimizer in run on this set. Next node is
discorvered and added to the images to optimize, optimize, ... and so on.


I find that this usually is a very robust way to produce a good initial
estimates. Since it doesn't consider the previous positions of the images
(except for the anchor image), it works in cases where the original
optimizer is stuck in local minima.

However, I found that the newer optimized usually works similary really well
when starting from y,p,r = 0 for all images.

But I have tested it only on a small amount of panos, and only one had more
than 20 images.

ciao
  Pablo