Re: Surface Rendering from Stereo-pairs

[Jonathan Joseph <jj21(at)cornell.edu>]

Larry,

While the theory behind photogrammetry (reconstructing the
3D topography from images) is, I believe, well understood, in
practice, it's really not that simple.

In the idealized case, if you have N images in each of
which you can see the same M landmarks (there are minimums
for N and M) then you can figure out the 3D locations of the
landmarks and the positions and orientations of the cameras - to
a factor of scale.

There are commercial products on the market which do this.
I remember looking into a few, but decided that they were
too specific for my needs.  Sorry I don't have any names
(it was a couple of years ago).

People that do GIS (forgive me but I'm not sure what that
stands for - maybe Geological Information Systems) or anyone
that wants to make a topographic map are, I believe, the largest
group of people to make use of photogrammetry.  They are
usually dealing with the special case of reconstructing topography
on what is basically flat ground with lumps, from aerial photographs.

Another set of users are those who want to reconstruct the
shape of objects without spending $$ on a 3D scanner.  Or
to reconstruct the shape of buildings.

If you want just the locations of a few points that you then
triangulate to form surfaces, that can generally achieved
pretty well if you have "good" landmarks - and images from
"good" vantage points.

The programs are basically triangulating to find the locations
of the points.  A landmark is generally not a precise point and
error is easily introduced in picking their locations.
A small amount of error in marking the location of a landmark
will be greatly magnified if the images are taken from vantage
points with a small angular separation.

If you want a very high resolution surface - then you won't
want to mark thousands of points in each image - you will
want something that uses pattern matching on a pair of images
to calculate control points for you.   Depending on the type
of image, this can work very well or very poorly.  Of course,
pattern matching will work best when images are very similar
(which is exactly when triangulation will magnify any errors).

I'm not sure exactly what your need is, but maybe I have
given you some leads on where to search.

The more variables you can remove from the equation (such as exact
location, orientation and optic paramaters of the camera for instance)
the easier it will become.

The more images you have, the more you can reduce the error (usually).
And the longer your processing time will be.

Good luck.

-Jonathan

Larry Busse wrote:

> Is anyone aware of a method or software package that is capable of doing
> surface mapping using a pair of stero images as input?  I realize it may
> be necessary to mark (or manually identify) common points in the
> images.  Any guidance or suggestions in this regard would be greatly
> appreciated.
>
> --
> Larry Busse                  Phone/Fax:   (606)344-1464
> LJB Development
> mailto:ljb@ljbdev.com       http://www.ljbdev.com

--

-Jonathan