T3D Re: stereo & computers

[roberts@xxxxxxxxxxxxxxxxx (John W Roberts)]

>Date: Mon, 13 Apr 1998 10:59:23 -0600
>From: john bercovitz <bercov@xxxxxxxxxxx>
>Subject: T3D stereo & computers

>Is it OK to ask a question on this list if it's for my job?  8-)

And is it a question for T3D or SD3D? :-)

>I've got a network of ropes attached to the inside of a thin-
>walled pressure vessel and to each other.  The ropes meet at what 
>I would call nodes some of which are on the vessel walls but most 
>of which are in space.  

If the ropes are attached to the *inside* of the vessel, then either the
nodes are also inside the vessel, or else the ropes pass through the walls
of the vessel.

Are the nodes *inside* the vessel, or *outside* the vessel? And if outside,
are they all around the vessel in 3-space, or are they all visible from some
single point? (And are the walls of the vessel transparent or not?)

>I would name the nodes A, B, C,....N.  The 
>connections between nodes I would name AB, AC, AF, etc.  There are 
>a ton of nodes and the structure already exists.  

>What if I took 
>pictures of it from more than one vantage point?  

Is that pictures from inside the vessel looking at ropes inside the vessel,
or from outside the vessel looking at ropes outside the vessel?

Are the position and direction the camera is pointed recorded for every shot,
or are you considering just some random snapshots?

>Is there a 
>computer program smart enough to give me the positions of the 
>nodes and the names of the connections from those pictures?

Just from the geometry issues involved, I doubt there's a computer program
available that could do the job *without human intervention*. A very simple
argument for this is that even with n pictures, there could still be
ambiguities - such as ropes that are hidden behind other ropes in every
photo, and possibly situations in which a rope can be seen but there is no
parallax available. There might be specific items for which the sensitivity
(in interpretation of position) to initial errors is much greater than for
other items.

If I were designing such a program, I would probably have it isolate out the
visible nodes and ropes from the first photo (or first two photos), and come
up with a way to describe uncertainties, ambiguities, and items yet to be
determined, then re-interpret and update this model with the addition of
each new view, the output of the program being a description of what's
reasonably certain, what's still uncertain, and perhaps a request to get
additional shots from (near) specific positions, to resolve remaining
ambiguity.

I believe I've heard of programs that do a reasonably good job on a 
3-dimensional *surface*, though even they can't be guaranteed not to
miss significant geometric details.

If you find any promising programs, please post.

John R


------------------------------