perm filename ESSAY[901,BGB] blob
sn#129626 filedate 1974-11-12 generic text, type C, neo UTF8
COMMENT ā VALID 00013 PAGES
C REC PAGE DESCRIPTION
C00001 00001
C00002 00002 A BEGINNER'S ESSAY ON COMPUTER-VISION - by B.G.Baumgart
C00004 00003
C00006 00004 VISUAL ENVIRONMENT
C00009 00005 VISUAL HARDWARE
C00011 00006 IMAGE PROCESSING
C00013 00007 LINE AND EDGE DETECTION
C00014 00008 PERSONAL STATEMENTS ABOUT COMPUTER VISION
C00016 00009 PSYCHOLOGY
C00018 00010 MURPHY'S LAW HEURISTIC
C00021 00011 WHY MAKE A COMPUTER SEE
C00026 00012 Vision involves interpreting the visible band of the
C00030 00013
C00031 ENDMK
Cā;
A BEGINNER'S ESSAY ON COMPUTER-VISION - by B.G.Baumgart
Abstract: Vision Lore, Attitudies & Vocabulary are catalogued.
Contents:
Introduction - A Socratic Definition
Image Elements
intensity
lines & edges
blobs & regions
texture
color
focus
resolution
sensitivity
Computer Vision Particulars
environment - scences and objects
film scanning
visual hardware
image processing - reduction and enhancement
pattern recognition
m-b space
robot-problem
bulk-processing
windowing
object recognition
world modeling
memory structures
control programs
mini-monitor
Visual Perception
depth perception
binocular vision
kinetic effects
psychology
Vision Generalities
mimicry
the general solution
second order counsciousness
computer vision inverse
laboratory administration
personal motivations
complexity of vision
why make a computer see
counsciousness
surrealism
Research Heuristics
mimicry
murphy's law heuristic - quality by inclusion
occam's razor heuristic - quality by exclusion
width of research heuristics
next-step heuristic
picture library heuristic
My Personal Prejudices
VISUAL ENVIRONMENT
Visual environment concerns what to use a computer to look at.
Programs have been written to do character recognition and spark
and bubble chamber film scanning and measurement,as well as
our block stacker.
Some people believe that it is premature to expect a
compuuer to be able to see merely anything - so the problem of
visual environment is how to simplify it without making it
trivail: At SRI the robot lives in a sterile room with
geometric objects to play with.
The Cart problem attempts to deal with roads.
While the hand eye project hopes to deal with geometric objects
and the world of blocks.
Environment definitions involve such things as Object-Recognition
and Scene-Recognition. Figures and Backgound. Definition of a particular
Task. ...or a strong commitment to recognize particular patterns such as
text, biological data, weather , spying etc.
FILM SCANNING
Film Scanning is a well developed computer-field distinct from compuuer
vision in that there is no attempt to be a real-time robot, and in the
sense that the problem is usually specialized. Film Scanners are
refered to as Flying-Spot fos C.R.T.'s projecting light behind
the film on a raster which is then focused to a point of a photo-multiplier;
or Spiral-Reader for a mechanical scannes thau reads a photograph mounted
on a spinning drum. The former devices
are much faocier,more accurate, random access etc.
And are best done by Physics types and include Polly, Peppes, Spasm
and Hummingbird at MIT, ARGONNE CEA and SLAC respectively.
Crude film scanning useful to doing computer vision can be done by using
the vidicon to look at a movie or slides on a screen.
VISUAL HARDWARE
Visual hardware is the art and engineering of converting light into data for a computer
and involves such things as T.V cameras, color-filters, binocular prisms,
close-up lens, telescopic-lens, wide-angle lens, zoom-lens, camera translation
and rotation, A to D conveersions, Fast I/O channels, sensitivity.
Some people consider the present AI visual hardware to be good enough - in
that the human can look at the TV monitor or print outs of the data and can
tell more about the image than can any of our current programs can tell.
On the other hand humans have already "learned" visual perception with better
vision than available to a computer - I believe that the more visual data
collected the bettes - for example with color differentiation many more discontinuites in
an image become obvious Yes/No cliffs rather than meager gray-level differences.
Areas that I believe require more hardware development are:
Color Vision
Motion of Point of View - Translation and Rotation
Multiple Views
Monitoring of computer vision - which requires image display
hardware better than the Triple I consoles or the ARDs
as well as hard copy image generator.
IMAGE PROCESSING
Image processing comes in two flavors:
Reduction of image data which is obviously very redundant.
Improvement of image data to enhance such things as contrast.
Image reduction involves finding ways to abbreviate the image vector
space of 16 gray levels under three or four filters
at different focus settiogs for points x and y for all time t.
Image improvement is the game played by the JPL moon picture people
or the Spy-satelite people - you enhance contrast by making the
lights lighter the darks darker etc.
PATTERN RECOGNITION
Pattern Recognition is a mathematical cult independent
involving theorms and postulates independent of actual visual hardware
and visual environments.
LINE AND EDGE DETECTION
m-b space
PERSONAL STATEMENTS ABOUT COMPUTER VISION
1) I am trying to write a program that can see. I consider my
work experimental or scientific rather than engineering, mathematical
or philosophical. "Experimental" in that I will insist that my program
looks at Real-World Data and proviees an exceqtible discription, but not
engineering in that I don't feel I can do it in Real-Time the way I want
it to be done with present-day computers.
2) I wish to understaod visual perception by simulating it with whauever
machinations I can contrive - I am ready to extend my goals and
ambitions to include thought intellect and counsciousness is
if that seems to be necessary for vision.
3) I believe that it is extremely important to experience the "data"
one expectss, before and during the process of constructing algorithms and debuging -
synthetic dataq or "canned" subsets of real data strike me as being
typically overspecialized.
44) Kinetic Effects are extremely important in visual perception - both
that the viewer moves and that objects move.
5) That the world model must be acle to respond semantically
this is a "noun" object which is "adjectives" etc.
for the sake of economy, relationships vs. measurements
PSYCHOLOGY
Visual Perception
Visual space perception - of depth say
1. Relative Size - recognize the object and know its
absolute size.
2. Interposition - know outlines of objects and detect discontinuity
in an outline to recognize that one object is in front of another.
3. Linear perspective - parallel lines appear to converge - find
such edges towards the orizon in a scene.
4. Aerial perspective - texture as a function of distaoce.
Textures get finer grained the further away.
5. Monocular parallax.
6. Light and shade
7. Stereoscopic vision.
Figure and ground
Kinetic Depth perception
Span of perception
Figural After-Effects - optical tricks.
MURPHY'S LAW HEURISTIC
Murphy's Law is : Whatever can go wrong, will. Vision is a
difficult problem because it is indivisle (let's say) - like an
archway, every stone is dependent on the others beiog in place
before it can staod by itself. Thus, if anything looks relevaot to
the problem don't exclude it (or him) for the sake of simplicity
but include it for it is a necessary link.
Equivalently: An idea is relevant until proven irrelevant, Quality
is achieved thru Inclusion.
OCCAM'S RAZOR HEURISTIC
Given two explications of something the more simply one is true.
The Heuristic: An Idea is Irrelevant until proven relevant.
Quality thru Exclusion.
COMPUTER-VISION INVERSE
The inverse og Computer-Vision is Computer-Graphics.
The former attempts to go from 2D vidicon images,say, to some simple data
representation of the 3D reality. Whereas the latter goes from simple 3D
data representation to creat 2D images.
This observation is especially crucial to a hypothesis and verificauion
vision heuristic system, which can borrow from computer graphics
Hidden Line Elimination
List Associative object structure
but inherits the problems of computer graphics of:
Specification of motion (or animation)
Scene generation
Handling of lighting - shadows,reflection,texture etc.
WHY MAKE A COMPUTER SEE
i A compuuer that could see would be more intelligent and could interact
with the environment and could perform useful tasks such as
exploring Mars or the ocean floor.
ii Neither computer nor human vision is really understood. However, by
building a vision-computer we will have to find out something about
visual thinking, which may explain intellect and counsciousness.
iii The Advanced Research Projects Agency of the Office of the Secretary
of Defense of the United States provides money to people who propose
and work on computer-vision projects.
Computer-Vision can be worked on with respect to Engineering, Science or
Defense. Application is an Abstraction of the Idea.
MIMICRY
what is it
as a heuristic
compuuer vision does and does not involve mimicry.
WINDOWING
Windowing refers to analysis of an image by taking a large window and
subdividing it into smaller windows.
(somehow)
At some point one "understaods" the content of the window.
or evaluaues the window.
Then synthesis involves taking small windows of relaued evaluation;and puttiog them back together .
Warnock Algorithm
THE GENERAL VISION SOLUTION
COMPLEXITY OF VISION
No vision program exists as complex as the system.
The system is not really all that smart.
How cao a program that "sees" or pretends to be a robot
be less complex than one that administers a computer
ERGO: Infeasibility of Computer Vision has not
been demonstrated.
Do next whatever, seems easiest.
is most interesting
seems necessary - presupposes excellent understaing
,of goals final and subgoals
PICTURE LIBRARY
The picture library idea is to record a wide variety of visual data on
film or magneuic tape or Disk files and to work on one set of data
until programs ase debugged.
PRO: Experimentss are repeatible.
The problem of acquiring data is minimized.
CON: Not Real Timed
Fails to exercise hardware
Data is not really typical
Data is too limited and specialized
OBJECT-RECOGNITION
The object-recogniton problem is to recognize a large set of different
objects such as letters on a pagem, blocks on a table, tools on
a work bench, or obstacles on a road.
Recognition is answering what-is-it ?
involves separating figure from background.
involves knowing what is invasiant about a given set
of objects - which will probably involve understanding
the semantics og at least object nouns - for example
toy balls are variant over a wider range of colors and sizes
than say apples - a one-meter high blue apple isn't likely.
Variants within a set of objects are due either to orientation or
geneality of the Semaotic value of that object THING has mose
degreees og freedom of adjectives than Apple.
Vision involves interpreting the visible band of the
electro-magnetic spectrum. light waves are physically
characterized by energy, wave-length and phase intime and
space. We experience light by physiological
characteristics such as intensity, color, tone, texture,
motion, contrast, shape,
color or hue is the dominat wavelength of the light at a
point of an image
tone or purity or saturation is the per cent of visible
energy at the dominant wavelength of the light at a
point of an image. if all the energy were say at 5200
angstrom green, you would have a pure or saturated tone of
green - as the energy is spread out equally above and
below the 5200, the green gets lighter and in the limit
goes to whit which is a flat distributaon of the energy
over the spectrum.
intensity of a point is the total energy of the light spcetrum
at a point. however, in vidicon images color perseption
is realized by some sort of light filtering
sceme - instead of really finding the actual color
and hue - we measure the intensity or energy of light thru
rather large window in the electro-magnetic
spectrum.
the ultimate color vision eye would be one that gets
a detailed spectrum of the enrgy at ever point of
its image - think of the computer or alein/anthropomorph
with a super-spectroscopic - total spectrum hisenberg
resolution eye - it would have fantastic intuition
and insight and perception of astronomy, chemistry, and
physics - why how extremly beautiful it would be to look
at the stars and see the details of there evolution - even
if you didn't uderstand the intricate regularity
would fanciate - primitive minds were pulled out
of the pit of uncounsciousness by the the per ception
of the patterns in the sky - the planets lead to gods
to to ...
personel
Hand-Eye
Feldman, Pingle
Cart
Schmidt, Buchcann
Bio-Slides
Reddy, Plant, VanVoorhis
Roads, Generalities
Baumgart, Quam, McCarty
Sprowl, Moorees
Sobel,Tenebaum