[photo-3d] Re: Algorithm Wanted

["Bruce Springsteen" <bsspringsteen@xxxxxxxxx>]

Robert Thorpe wrote:

> If you are looking to correct for rotational error, I would find two
> identical points in each picture. Hopefully one on the right edge
> and one on the left edge. I would rotate one of the pictures till it
> looked right vertically and tape it down. I would align the higher 
> of two points in the second picture with the same point in the
> first, taped down picture, think of putting a pushpin through the
> pictures at that point. Then I would rotate the second picture
> around the push pin till the other two points were aligned. Trim
> the tops and bottoms so they match and then correct for the
> window.
> 
> I suppose this isn't anything that you were interested in and that
> I have barked completely up the wrong tree.

This is pretty much the right tree exactly - in fact I've had visions
of pushpins in my head as well.  I'm trying to avoid rotational
error in the first place, more than correct for it.  I'm not however
assuming that we have usable edges, or that we know what edge
is left or right - that's part of what I want the algorithm to do, to
help us establish left/right/up/down when there are no straight
edges or we don't trust the ones we have.  Also assume there
are no clues in the subject matter, no way to just look at the picture
and make a guess at rightside up.  Maybe it's an overhead shot
looking down on a crowd of people - no horizon, no indication in
the subject of what orientation the stereo camera was in.  Just a
note on the back of each print, one saying "left image" the other
saying "right image".  Is Robert's algorithm foolproof in THAT
scenario?  Is it the most efficient or practical solution?  I don't
know, but I suspect there is a problem.   Am I wrong?

Bruce