Algebra For Stereo-Scopic Image Generation

Algebra For Stereo-Scopic Image Generation

Post by Wade Bick » Tue, 02 Jan 1990 19:52:00

        Thanks Leo for pointing out this discussion in amiga/comp-amiga.
     I did not have direct access to this discussion and didn't know it
     even existed.

        First, I'd like to mention that Haitex will be releasing 3-D
     glasses for the Amiga within about a month.  I like our glasses
     better than the Sega glasses because they have bigger lenses, and the
     lenses are on a visor rather than "glasses" (kinda like a welding visor)
     which is more comfortable, especially if you already wear glasses.
     This is one of the things I've been working on for the past few months.

        The following is a letter I sent to the makers of a ray trace progam
     which will directly support the glasses.  I hope this helps those of
     you working on 3-D projects.

        Note for Leo:   This entire methodology was developed by looking at
     the point of a pencil held between myself and a window, and looking at
     a branch outside the window.  This methodology is totally intuitive, and
     avoids the need to perform any rotions to acheive the parallax effect.
     Furthermore, I have been able to bring object a significant distance
     out of the screen with no terrible ugliness.  I'd like to talk to you
     about all this in more detail so why don't you send me your ph# or
     call me [(619) 421-5602] and I'll call you right back so I pay for the

FROM:      Wade W. Bickel, Haitex Resources.            DATE: 12-10-88
SUBJECT:   Parallax Rendering Algebra.

       In this letter I will try to go through the mathematics involved in
    creating a parallax stereo pair.  I am not a mathematician so please
    excuse me if my teminology is non standard.

         For computational simplicity I perform the following rather convoluted
      system for generating a stereo display.  The idea is that there is
      an imaginary 3-D grid which cannot be re-positioned, and points to
      be translated onto a region (which represents the screen), both of
      which can be freely re-positioned.  If you picture the display as
      a window we are electing to move the viewer's head, the window, and
      the objects outside that window, rather than re-align the axes.
        The logic can be decomposed into a system of moving axes if you wish,
      but this system eliminates the need to compute the length of some
      projections using the distance formulae and therefore is the one that
      I utilize.

             1)     Points to be translated are represented in cartisian
                 (x-y-z) coordinates.

             2)     The viewpoint is assumed to be centered at the origin
                 and facing in the direction of positive z.  To visualize
                 this, you the viewer, are at the origin, the direction
                 you're facing is positive z, and negative values of x lie
                 to your left.

             3)     A plane lies perpendicular to the z-axis (ie: parallel
                 to the x-y plane) at some distance in the positive z
                 direction.  Within this plane is a rectangular region, not
                 quite centered about the z-axis, and aligned with the x and
                 y axes, which represents the screen.  For our purposes we
                 will call this region either the "left-screen" or the
                 "right-screen", or just "screen" for the general case.  
                    The screen is offset with respect to x depending on which
                 view (left or right) is being rendered.  For the left eye
                 the screen is offset to the right.  Thus if one were to
                 display the point [0,0,zP] (ie: where the z-axis intersects
                 the region) for the left-screen the point would appear to
                 the left of the center of the actual display bitmap.  For
                 the right eye the region is offset to the left.  These
                 offset's will be refered to as the "lefteye-offset" and
                 "righteye-offset" and are calculated to be one half the
                 distance between the viewer's eyes, in whatever units of
                 measure have meaning to the programmer/user.

             4)     Points to be translated are assumed to have a positive
                 z coordinate; in other words, you should clip with respect
                 to z before trying to translate a point.  Clipping with
                 respect to x-y should probably be done after translation.
                    When translating a point for the left eye, that point
                 should be shifted right by twice the amount the screen
                 region was shifted to the right (ie: the total distance
                 between the eyes).  Lets call this the "EyeSpacing". The
                 net effect of the shift of the screen and the point is
                 to properly create a perspective.

         Given these conditions and relationships, to translate a point onto
      the screen the following algebra is used:

            The point to be translated (always positive z) and the view point
         (always [0,0,0]) define a line.  These two points also form a right
         triangle, call it triangle "A", with the z-axis.  The line also
         passes through the plane in which the screen (as previously defined)
         lies, and this intersection is the one we wish to find.  Note that a
         triangle is formed between this unknown point, the viewpoint, and
         the z-axis, call it triangle "B".

            Triangle A and B are similar right triangles.  Since we know
         the z components of both triangles, and everything about triangle
         A, we can calculate the disired intersection as follows:


                   VP = the viewpoint, [0,0,0];
                   P  = the point to be translated (+z);
                   S  = the plane on which the screen lies (perp. to z);
                   Q  = point on the line defined by VP and P, and on
                          plane S.

              the following relationships are true  

                   Qx/Qz = Px/Pz;  -->  Qx = (Px * Qz)/Pz;    

                   Qy/Qz = Py/Pz;  -->  Qy = (Py * Qz)/Pz;

                                        (note avoidance of distance fomulae
                                            calculations because VP = [0,0,0])

              Where Qx and Qy are the respective X and Y coordinates to
           possibly be displayed.  This two-d-point should be checked to
           see if it is within the "screen" (as previously described) and
           if so it should be displayed.

              Implementation of this is not nearly as difficult as all this
           implies.  Simply do the following.

                        1) for each view add or subract the constant
                             "EyeSpacing" to the x-component of points
                             to be translated.  Hopefully this can be done
                             at a very low level of the point's position
                             calculation.  (Add for left eye view, subtract
                             for right eye view).

                        2) translate the point as described earlier.  Qx and
                             Qy represent the actual translation to the
                             plane on which the screen lies.  At this point
                             determine if this point lies in the appropriate
                             "screen".  [Remember that the screen is a region
                             which is shifted left or right by half the
                             EyeSpacing (or: left/right-offset)]. If so then
                             the point should be rendered.

                        3) When rendering the points must be recentered to
                             account for the fact that [0,0] on the display
                             is the upper left corner, and we want it to
                             be the psuedo screen center.  At this point,
                             adjust the x compensation by one half the
                             EyeSpaceing.  (move the center left for the
                             left eye, right for the right eye).  This should
                             be an integer operation.

              Implementation of this is not nearly as difficult as all this
           implies.  Simply do the following.

              I hope I've not gotten anything backwards.  Sorry I've made it
           seem overly complicated, but I tend to think of this in pictures,
           not language.


              Now on to my question:  Is it not worthwhile to think of objects
           in polar coordinates?   It seems to me than much of the math
           required to do the real-time calculations necessary to provide
           rotation of both the objects and the viewpoint(s) would be
           allieviated by defining object this way as opposed to cartesian
           coordinates.  I am about to give up on this approach due to some
           flaw in my calculations, which took me about a week to work out,
           and translations in cartesian coords are well documented.  Seems
           a shame though as it should reduce the needed calculations to less
           than half.

              If anyone would care to discuss the topic I might give it one
           more shot.   Should I post here or keep this E-mail?

              All (civil) comments are welcome.


                                                             Wade W. Bickel
                                                             Haitex Resources.

        PS:   In order to get to this news-group I have to utilize the
            Unix side of the local node.  Not knowing my way around in
            here yet, would someone please mail me the Stereo-vision
            topics found here for a while (and perhaps post letters for
            me as well, which I would send via mail)?

        PPS:  Sandra (?? I think that was the name of the original poster),
            If you'd like I could mail you some Modula-2 examples to perform
            the translations.  Of course, if you want controlled rotation,
            that's what I'm working on now so you'd have to wait.


1. Stereo Pair generation

        Be forewarned, Rayshade uses the Rotational model for stereo
pair generation which gives inaccurate results.  You can use a dual
eyepoint model by either 1) hacking rayshade which has been done here,
or 2) figuring out where you want each eye to be and just specifying
each eyepoint seperately to rayshade.  Either way, rayshade does twice
the work as for a single eyepoint.  To save work, read:
        Simultaneous Generation of Stereoscopic Views by Adelson,
Bentley, Chong, Hodges, and Winograd, to appear in COMPUTER GRAPHICS
FORUM 10(1) 1991.


        Ray Tracing Stereoscopic Images by Adelson & Hodges, submitted

or contact Steve Adelson at the College of Computing at Georgia
Institute of Technology.



- 919-737-7279, NC State U., Computer Sci. Dept., Box 8206 Raleigh, NC 27695 -


3. Web page on mesh generation / grid generation: New release

4. Question to French Raytracers

5. BVS Reviews February 2001 Edition

6. Tree and plant image generation

7. "swish" software help

8. Plant and re tree image generation

9. Plant and tree image generation

10. Software for image generation from 3d data?

11. Tree and plant image generation

12. slow image generation