[Visualpython-users] Anaglyph stereo rendering in VPython -- could it be improved?

[Stanley M S. <ove...@ea...>]

Hi, VPythoners,

I'm trying to get nice looking stereo anaglyphs of VPython models, but
the built-in anaglyph rendering seems to have some problems for my eyes,
and maybe everyone else's eyes, too.   I'm no expert in anaglyphs but
have been learning lately on various Internet sites.  I bought some good
quality red/cyan eyeglasses (with diopter compensation in the red lens)
and a clip-on red/blue pair that I clip onto the red/cyan glasses and
flip up or down to remove or use them as necessary for particular
anaglyphs.   I can also clip the red/blue set onto my regular glasses to
use alone.  When I set some of the VPython demos to render in red/cyan
or red/blue anaglyphs (for example, "scene.stereo = 'redblue' "), the
rendering does produce anaglyph displays but I have pretty bad
"ghosting" problems.  That is, my eyes and brain can merge the left and
right eye images into a stereo virtual scene only in a portion of the
visual depth.   A significant part of the rendering has too much
horizontal separation between the red/cyan or red/blue pair of images,
which is beyond my capability of merging visually.  I can see the red
and blue eye images separated rather than merged.  This is "ghosting."
I wonder if other users have the same problem.

I have seen many examples of good and bad anaglyphs on the web.  Look at 
this red/blue anaglyph image: 
http://www.dogfeathers.com/3d/povray/3dgdddd.html .
Even though this is a synthetic object (that is, constructed by a 
program, not an actual camera photograph),  it has excellent 3d and 
little or no ghosting to my  eyes, when viewed with red/blue glasses. 
It came from this site, which has more excellent examples:
http://www.dogfeathers.com/3d/povray/index.html .  I haven't been able 
to get such good, clean 3d from the VPython demos.

Here's another example of remarkably good anaglyphs from computed
objects:  http://www.dogfeathers.com/java/ccppoly.html .  Set the Java
applet's stereo mode to "anaglyph" and its color scheme to
"red-blue-black", put on your red/blue eyeglasses, and set the
"sequence" to the left side of the slider, then work your way up through
the first few "sequence" settings for an eye-popping experience.   You
can play with image separation by adjusting a slider on the applet.
It's also an experience to adjust the "roundness".   This display looks
best to me when I use both of my anaglyph eyeglasses together (red/cyan
with red/blue also).   This reduces the ghosting to almost zero, but it
does make a rather dark image.

Here's a site with astonishingly good anaglyphs that were created either
from stereo pair photographs or from ordinary flat photographs that were
carefully, artistically, retouched into stereo using the software
created by the web site's Italian owner.   This is a "must-see" if you
have stereo anaglyph eyeglasses:  http://digilander.libero.it/kems/  .
If you don't have anaglyph stereo viewing eyeglasses, it's worth
searching the Internet for stereo sites to buy a set (such as
http://www.berezin.com/3d/ where I bought mine).

Here are some examples of various anaglyph rendering transformations,
which show full-color, half-color, greyscale, etc. comparisons.
http://www.3dtv.at/Knowhow/AnaglyphComparison_en.aspx  You'll see that
some have a lot of ghosting while others give up some or all of the
color information in order to reduce or eliminate ghosting.  The
anaglyph color separation in vpython is done strictly by masking the
color channels, so they are all or none, with no attempt to blend them
the way that the matrix transformation in the comparison web site does.

I think one problem with VPython's anaglyph rendering is that the eye
separation attribute is hardcoded to be 1/30 of the focal length of the
camera lens. (Actually, the comment in the VPython code says 1/30 but
the actual division is 1/60, but maybe that's a result of some other
factor of 1/2 in there -- the math for the focal length isn't clear to
me.)   Some people's vision need less eye separation in the anaglyph,
resulting in less differential between the left eye and right eye images
in the superimposed rendering.  The 1/30 rule comes from general
experience with cameras.   Most people can see merged stereo if the
distance between the two cameras is 1/30 of the distance from the
cameras to the nearest object in the scene.  That's not the same as 1/30
of the focal length of the camera lens, of course.   (But maybe due to
the way that vpython computes its camera "focal length," the result may
be equivalent to the photographer's 1/30 rule.  It does not seem
equivalent to me, though.)  However, in macro (close-up) stereo
photography the eye separation needs to be reduced and in landscape
photography it is greatly increased to obtain a good stereo effect.

I think VPython should expose the eye separation parameter (1/30 or
whatever) as an attribute of the Python "scene" object, that is, so you
can set the scene eye separation by "scene.eyeseparation" just as you
can set "scene.stereodepth."  That way the display can be customized by
the user to whatever eye separation works for his eyes.  It doesn't seem
like a difficult change to make, since it would be quite similar to the
other settable attributes of the scene.  It's just beyond my programming
skills to make the change to the C++ code, otherwise I would try it
before suggesting it here.

My point in all of this is that I think VPython's anaglyph rendering
could be better with a few changes.  Are there any comments from the
anaglyph experts in the group?

Stan Sokolow