README for synthetic stereo satellite images of earth   10 May 92

Images:  
earth.l.gif, earth.r.gif        stereopair
earth.gif                       raw data

The images began as a 720 x 360 file of floating point elevations for the
planet (found on hanauma.stanford.edu).  I converted these to bytes to
obtain a range image, a greyscale image. I used the "rawtopgm" program to
convert a headerless byte-file into .pgm format that can be read by my
viewing program. I view the images stereoscopically by running two copies of
the viewing program (I use "xv").

Next I fed the range image and the elevation data to a program that added
disparity info to the image.  The disparity could be added by shifting the
actual texture (the range image); or by superimposing a texture and shifting
that.  Various textures could be used.

(I used a texture with no pixel-by-pixel correlation.  This was simple and
yielded the smallest "grain".  Several issues arise: should one use a
constant-amplitude speckle noise, or should the noise amplitude vary with
the pixel level (like quantum noise)?  The problem here is that a constant
amplitude will be large for dark pixels and unnoticable for bright pixels.
Also, the dark pixels aren't seen.  Using a scheme where speckle is added to
small-valued pixels and speckle is subtracted from high-value pixels yields
a form of gain control.  This attentuates edges that are useful for stereo.)

Each method for adding disparity has its disadvantages.  The problem with
adding textures is that they degrade the greyscale image; the problem with
shifting the images is filling in the resulting gaps, and a possible lack of
features to match up for correspondence.

The images here have a pixel-disparity for every 250 meters of elevation
difference.  The noise has value -40, 0, or 40 greylevels.

It takes quite a while to achieve stereopsis for these images, but it can be
done.  The ocean floors and midoceanic ridges are particularly good.  (Most of the relief of the earth is in the oceans.)  The himalayas are probably
too extreme to be fused, other ranges can be seen in depth.

-David Honig

PS: The white artifacts are in the original data set.

.