Re: Computer Compositing

[T3D Larry Berlin <lberlin@xxxxxxxxx>]

>> With today's tools you could put those images
>> into your computer and *fix* the parallax problem in the images themselves
>> then reconvert to film. This would be a much more ideal solution if the
>> problem is severe enough and worth being fixed. 

> Have you actually tried it?  

Yes, I've done it. One image currently on my 3DZine site (Lake
Mendocino) used these types of manipulations but I wasn't trying to correct
parallax for projection. I had a stereo image taken with a single camera and
the water moved while I repositioned for the second shot. It's a
hyper-stereo taken with a 70 mm lens. The image had two problems. Due to the
angle of the sun on the lens, there were annoying reflections in each image
but in different places. I carefully, one small area at a time, painted over
the reflection with image information from the other side. This required
that each snippet be reshaped to a similar stereo-disparity as the image
behind the reflection. The other problem was the water movement. I used both
images as source material and reshaped the information to match that of the
correct 3D projection of the surface of the water. The result is as if I
took both images at the same moment. Similar tricks and distortions,
carefully applied could reshape the parallax in a given image until it is
within a desired specification. (maybe not any image, ... some are pretty
complex)

> The image is scanned into a perfectly planar
> bitmap and "objects" as such do not exist.  

With any stereo-pair image you have more than a planar
object. You have more than a wire frame and more than an object with painted
polygons. The actual structure as well as a great amount of detail is
recorded as parallax between two flat bitmapped images.

 Just like a collection of polygons provide a semblance of rounded surfaces,
little snippets of image, distorted horizontally in subtle ways, and traded
back and forth between paired images, will aproximate the plane
perpendicular to any curved surface. Do this in sufficient detail and the
image material makes the work indistinguishable. No more effort than a
detailed painting and easier in the sense that you have more control of your
*paint* and can easily undue mistakes and repeat your effort. In this case
the source image itself is a photograph so the result is a synthesis of
recorded reality instead of a *painting*.

>With a great deal of effort you
>can mask outline an object and cut it to its own layer, but then you must
>re-create what was hiding behind the object, 

Only certain edges. There are usually desirable pixels handy nearby.

sometimes you can borrow
>(clone) from the other image.  That works, though is extremely painstaking,
>for flat objects; but an object viewed obliquely already has a continuously
>varying parallax change, which you decide to alter subtly!  Good luck!

Learn to think in 3D by working in 3D. It's a process of
tweak and pull until it's right. (Working with parallax is working with
horizontal distortions and shifts) Accept that you can't do the whole thing
in one step and start taking lots of small steps.

>If you START with a computer generated image, where every object is either
>coded on a layer, or as an object, or exists in virtual space (ie,
>raytracing), then it is easy to adjust the various parameters of parallax
>and perspective.

That's true and it helps to know how to correct problems
from many directions. It's also useful to be able to incorporate stereo
photography into work that is computer oriented. It's far more realistic
than painted polygons and can be done in the 3D rendering process.

>One friend of mine, B.D., who has been making stereo photographs since
>before I was born, supposed that all one needed to do to turn a flattie into
>stereo, would be to scan a photograph, rotate it a degree or two, scan it
>again and rotate it the other way a degree or two.  I explained that the
>result would look like a perfectly flat photograph that had been curled
>around a vertical axis.  Each eye would see exactly the same information,
>however, a distortion field would be present that would cause curvature of
>the plane.

Your friend was on the right track except that you have to
do that turning process for each part of the image independently. AND you're
right, except it would be a flat (not curled, but maybe this is what you
meant...) reproduction plane not perpendicular to the viewer. The basic fact
is that you are supposed to see the same image elements with each eye. If
you don't you encounter retinal rivalry. Parallax is only small amounts of
horizontal displacement. Depth is all relative to the repetition spacing of
each picture element. Compression or expansion in a horizontal direction
accomplishes the needed parallax for any given small area. (NOT a one-step
direct solution for excess parallax overall.)

>Similarly, he supposed that with a computer, all one need do is click on the
>apple in the photograph and move it so its perspective would not be so
>great, for example.  But the computer does not "see" the apple, though if it
>is strongly marked against the background you can automatically wrap a
>selector mask around it, and move it -- but that exposes white underneath!
>Yes, it can be done, but it is really difficult.

Basically your friend had the right idea. If he had the
chance to get into a computer and start exploring his ideas they would have
refined to reality in various ways. To make the apple round you first copy
the image so you have a pair. Then you take layers of the apple and move
them sidewards one pixel at a time, and layer by feathered layer. Watch the
operation in stereo as you move it and it's like doing sculpture. It may
take several attempts at first but after awhile you can figure out shortcut
methods.

The thing that leaves a white area behind is a totally unnecessary
limitation common to some graphics programs. BUT there is a workaround.
Simply start by pasting two or more copies of your image where you can
freely grab areas you want and not care what's left behind. Then move your
selection to a work image that isn't disturbed. 

Better yet, find a copy of Aldus PhotoStyler 2.0 SE (now considered obsolete
but one of the best I've found!!!), or PhotoShop or a number of other photo
kind of programs that allow lifting your selection off the page (float)
without changing the actual image that was there. Any program purporting to
do graphics that doesn't allow this IS NOT WORTH BUYING!!! Note that
different programs use different terminology so sometimes it takes a while
of use to figure out whether or not an operation is supported. Generally
many programs offer selection tools, and fewer of these allow you to really
manipulate the selections comprehensively. It seems this is because the
authors either couldn't figure out how to do this or they didn't see any use
for the utility or they haven't researched existing features in older
programs carefully enough. For stereo work it's THE ESSENTIAL INGREDIENT. I
use it all the time in 2D work too so I feel it should be standard in all
graphics programs but be warned, it's not always there.

If you're at all into photo images and high resolution results you should
plan on something like Photoshop and a computer optimized to work with
larger images. Then all or most of these tools are available and function
with optimum results. 

Have fun,
Larry Berlin


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