Re: Computer 3-goD (long)

[P3D John W Roberts <roberts@xxxxxxxxxxxxxxxxx>]

>Date: Thu, 31 Oct 1996 15:09:09 -0600
>From: "P3D Gregory J. Wageman" <gjw@xxxxxxxxxx>
>Subject: Re:  Computer 3-goD

>John Roberts replies:
>>>Translating a single 2D image into a 3D image requires the computer to
>>>recognize the *content* of the image 

>>Not necessarily - that's how a human would do it, but machines aren't
>>constrained to use the same techniques humans do. A photograph contains
>>information that a human viewer doesn't necessarily use.

>I disagree.  The human viewer has a HUGE repository of knowlege and
>processing capacity including shape recognition and the ability to
>"fill in" holes in the information.  You can cut off Aunt Tilly's
>head at the neck, and you would still recognize that it's Aunt Tilly.
>Traditional machine-based pattern matching would have a difficult
>time dealing with that.

I agree with what you say, but that wasn't my point. There may be ways
completely different from traditional pattern-matching to extract some
depth information. For instance, the degree of focus may vary with the
distance of the objects in the field of view, and if the computer program
can quantify this well enough, it can get some relative depth information.
(And by Peter Abrahams' recent description of chromatic aberration as a cue
to accommodation, if the camera's lens had sufficient chromatic aberration,
the uncertainty of whether the object is in front of or behind the point of
focus may be resolved.) That may not be sufficient in itself, but if a
number of "tricks" can be used together, it may be possible to get a
surprising amount of depth information (in many, not all photos) without
having to recognize Aunt Tilly. (I'm not saying that it can certainly be
done - I'm saying that it's "not necessarily" the case that the techniques
humans would most readily use *have to* be what the computer program uses.)

>For example, we have a very strong visual recognition system for faces.
>It is so strong in fact, that we see faces when they aren't even really
>there...

Another example besides the one you gave - it's extremely difficult to see
a pseudoscopic view of a human face as being inside out. By intense
concentration it's often possible to override some of these automatic
perception mechanisms - an interesting phenomenon in itself.

>Our visual cortex has processed
>the image so heavily before we consciously "see" it and we aren't even
>aware of how much it is doing.  This automatic processing is what makes
>us subject to all of those optical illusions that make straight lines
>seem bent, and so on.

It's amazing how many different techniques are used in human visual processing.
There was a really good television program a year or two ago that described
some of the neural structures that perform edge detection, angle detection,
and so on. Considerable processing also goes on before the visual information
even reaches the brain. I don't know how many different processes the human
visual system uses, but I wouldn't be surprised if it's hundreds. Many (all?)
optical illusions are cases in which the processing techniques break down -
something comes along that they weren't designed to handle.

>Humans are capable of recognizing objects that are out-of-focus and
>obscured.  Ever watch a badly snowy television picture and still make
>sense out of it?  Try that with a computer!

A lot of work is being done in this area, including research with neural
nets and photonic processing. Some of the accomplishments described in the
literature are really amazing - recognition with changes in scale, position
and rotation, recognition based on just a few dozen pixels in a visually
"noisy" environment, and so on. As you indicated, much of this is for fairly
specialized applications for now. Even fairly straightforward techniques
such as "speckle interferometry" can be useful in many cases.

>>It would be more productive to try to avoid phrases like "a computer knows
>>this" or "a computer does that" - though I'm sure you meant it as a shorthand
>>expression, it can be confusing to people who don't know much about computers.

>Thanks for the lesson in computers.  I've only been paid to program them
>for the last 17 years.  

>I didn't use the phrases you ascribe to me-- 

You used the following phrases:

  >Translating a single 2D image into a 3D image requires the computer to
  >recognize the *content* ...

  >I'm not sure at  what age a human being acquires these
  >skills, but you're talking about a computer with at least the
  >cognitive abilities of an older child! 

*Computer recognize*, and "computer with cognitive abilities". I didn't
say *you* didn't understand the distinction, I said it could be confusing
(I should have said "inadvertently misleading") to people who don't
understand computers well. Computers being so important in our society,
it's a shame that so many people have significant misunderstanding of what
computers are and what programs can do (a classic example is the human
reaction to the "psychotherapy" program). This is relevant to the discussion
of the ability of computers to process stereo images because if people get
into the habit of thinking that the hardware limits what the computer can
do, then they tend to think that computer hardware built in the future will
be able to perform computational tasks that are inherently impossible on
current computer hardware (storage space and execution time aside).

I'm probably being sloppy in nomenclature myself - would it be more accurate
to say that the process(es) perform the instructions implemented in the
program, and therefore implement the algorithm? :-)

>I said that the *human* recognizes which shapes are part of one object
>and not part of another object, even when they overlap, which a computer
>cannot successfully do without the same sort of reasoning process occurring.

What do you mean by "reasoning"? Do Mathcad or Macsyma employ "reasoning" when
they provide an algebraic expression as a solution to an indefinite integral?

>>While the research continues, I am not
>>aware of any convincing evidence presently available that human thought and
>>perception are *inherently* indescribable.

>Since we humans cannot even yet explain the process by which we "think", and
>as far as I know no one has yet proven whether or not the "mind" is simply
>a product of brain function, as opposed to something more metaphysical,
>I don't see how you are anyone can confidently predict that we will ever
>be able to make a machine think like we do.  

I don't either, and so I don't make that prediction. On the other hand,
I'm not aware that any "showstoppers" have been encountered. This is a
controversial topic, with strong opinions on all sides. There was a useful
article regarding the debate in Scientific American several years ago.

>As you yourself pointed out,
>a process must be described to the computer unambiguously and in the
>smallest detail.

It has to be made available *once*. And the description can be hierarchical
and as complex as human ingenuity can devise, including some ability to make
arbitrary decisions and to resolve ambiguities. This can potentially be
useful both for image interpretation and for creating realistic stereo 
graphics.

>Yamaha, Korg and other musical instrument manufacturers have just begun
>shipping a new generation of instruments based on just such a physical
>modelling algorithm, using a host of special-purpose Digital Signal
>Processors.  

Thanks - I wasn't aware of that.

John R


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