Re:On-line analysis of your stereo photographs...

[P3D apec <apec@xxxxxxxxxxxxxx>]

Just curious to know what the "cyclopean view" is, found by Jacques.
Following is a reply to my questions from Institute for Theoretical
Neurophysics.

Albert Lin
----------------------------------------------------------------------------
----------------------------
Sender: rdh@xxxxxxxxxxxxxxxxxxxxxxxxx
Date: Wed, 12 Mar 1997 10:01:34 +0100
From: P3D Rolf Henkel  <henkel@xxxxxxxxxxxxxxxxxxxx>
Organization: Institute for Theoretical Neurophysics
To: apec <apec@xxxxxxxxxxxxxx>
Subject: Re: Limitations/Cyclopean view
References: <1.5.4.32.19960312042913.00687ddc@xxxxxxxxxxxxxx>

Dear Albert Lin,

first of all - thanks for your comments.

Here some thoughts about your questions:

> 1. Is there any limitation for stereo images?

Yes. Basically, my algorithm trys to model human stereo vision, so the
cameras should not be spaced apart very large. This is in contrast to  -
for example - feature-based stereo algorithms, which can handle very
large distortions under some restrictive settings (The outline of the
objects isn't allowed to change between the views -> basically only 
cylindrical objects allowed). However, there *is* a method to combined
several stereo images (i.e. a movie) into a three-dimensional scene.
This is currently a research topic, and might end up in commerical
product (hopefully :-). The idea is again based on modeling human
perception. We'll see if it works out.

> For example, we shot a pair of left/right images at different position and
> viewing angles respectively. 

If you make your stereo images available, I might try to process them.
Currently, the following limitations apply:
a) the images have to be aligned properly.
b) humans should be able to fuse the images without difficulty (than,
there's a good chance that my algo can do as well).
c) without camera calibration, only relative distance information is
recovered.

But your can check out some VRLM-1.0 examples of what can be done with
two stereo images at the URL 

http://axon.physik.uni-bremen.de/~rdh/research/stereo/vrml/

> 2. Does cyclopean view works with only one 2D image? 

No. First, the cyclopean view is a fusion of the left and right images.
Second, it is possible to generate an image from *every* position
between the original camera position when the stereo images were taken.
Somewhat hidden in the Cyclopean View-Page, there is a link to a small
movie showing this (at the bottom).

>Is it possible to convert single 2D image to 3D image?

No. Not with these methods. I surely belief that this will be possible
(to some extent) sometime in the future (keywords are
shape-from-shading, shape-from-texture, etc.), but all methods known to
me do not work very well right in the moment. The only thing I know of
are programs were you, the user, is specifying the 3d-structure of the
scene. The 2d image is just mapped onto this structure.
 
> 3. How much distortion will be generated by cyclopean view?

Well, with view-angles corresponding to human geometry, not very much.
If the view-angles are too large, some image areas in the left view
won't match the right view and vice versa. For these image areas, the
cyclopean view can only be guessed. 

Hope this answers some of your questions. I had a quick look at your
website, seems to be very interesting. However, at the time of this
writing (8.45 am in Germany), the link is very slow. Might check in at
some later time. 

Anyway, again, thank you very much for your interest. 

Rolf Henkel

Institute of Theoretical Neurophysics, University Bremen, Germany
-----------------------------------------------------------------
Email:  henkel@xxxxxxxxxxxxxxxxxxxxxxxxx
URL:    http://axon.physik.uni-bremen.de/~rdh/research
-----------------------------------------------------------------



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