COMMENT ⊗   VALID 00003 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	Possible discussion topics
C00004 00003	1: Is The Low-Level defined? (notes by TOB [ and DBA])
C00014 ENDMK
C⊗;
Possible discussion topics

TOB 10 Oct
1. In a biological system what are good places to start using
	context?  Computer system?
2. What can be parallel and what can be serial?
3. What has been accomplished?
4. What will be accomplished?
	5 years
	10 years
5. Semantics: specific visual examples?
6. How to do x?
7. What is a diagram of a good vision system within the
	state of the art?
8. What incremental improvements are possible for current systems?
9. Is the notion low-level defined?
10.Is the notion high-level defined?

DBA 18 Oct
These are papers we could discuss
11. Arthur's thesis proposal
12. Marty Tenenbaum's recent SRI memo
13. Feldman/Yakimovsky on decision theory in vision
14. Max Clowes' notes for his book
15. Recent MIT vision stuff - Minsky's frames, Freuder's advice
16. Recent MIT vision stuff - Marr and Horn

1: Is The Low-Level defined? (notes by TOB [ and DBA])

There has been a strong emphasis in AI on knowledge and using knowledge
for finding features such as edges.  Is the term low-level even defined?
Two concrete questions:
	Are there useful features of a picture which can be calculated
	in a context-free way?
	Is there a parallel way of calculation?
The revolutionary progress of semiconductor hardware raises significant
questions; we cannot now say what would be useful on the basis of
simulation of parallel algorithms.  Many of our considerations have
been based on "current machines are almost fast enough.  What we need
are clever programs which do not do much low-level work".  
There has clearly been an emphasis on optimal decision structures which
do no unnecessary work.  Optimality has a very different sense when
parallel hardware is cheap, and for systems with many goals and a
complex universe.  There is not a good analogy we feel, with current
examples of goal-oriented systems which have very limited goals,
limited abilities, and exist in a simple universe.

We will attempt to answer the question by examples of systems which
do appear to be low-level in biological vision and which could be
low-level in computer vision.

1.  Lateral inhibition--Retinex
A word of explanation.  This system eliminates the effects of smooth
gradients of lighting (including spectral variation) for color vision
and edge detection operation.  Three phenomena are relevant here:
color constancy, brightness constancy, and the Cornsweet effect.
The last is strong experimental evidence that the human visual system
does not find edges at large gradient of intensity, but only discontinuity
in intensity or its gradient.  The teleological advantage is clear:
reflections and curved surfaces are common in nature, and using a gradient
operator as is usual in computer vision would give us whole surfaces of
spurious edge points.  Lateral inhibition has been used in the edge finder
used in the MIT copy demo (Binford), and the papers by Horn and Marr
are recommended.  [ Caveat by DBA.  This system is not magical ie it can
be fooled.  So while retinexing may well be a good thing to do, it is
essential that any system using it be able to override it/modify its
interpretation of it if conflicting evidence is received.  The same
goes for stereo depth maps, or at least those based on features.]

Since the system appears to be implemented in the retina, and appears
not to be sensitive to gradients, there is a question of how people
see gradients in surface apparent brightness, which are used for
guessing curvature.  
	There may be a non-retinex channel, particularly in the fovea
	which may have some one-one connections to the ganglion.
	They may be inferrable from the Retinex process.  Consider a
	non-uniform surface on a uniform surface.  When we perform the
	Retinex operation, the contrasts at the various portions of the
	boundary are different.  When the inverse computation is performed
	the resulting signal is non-uniform, because of the variation
	along the boundary.  Thus, a non-uniform reflectivity does show
	up while a non-uniform lighting does not show up, just because
	the reflectivity change manifests itself at discontinuities,
	while lighting changes don't change values along discontinuities.


2.  Hubel-Wiesel cells----edge operators
There is good reason to think that the Hubel-Wiesel cells are edge
operators; there are questions about the other assumptions of their
study, that these cells are innate and that there is a hierarchy of simple,
complex and hypercomplex.
There are several examples of such edge operators: the Hueckel operator
evaluates terms through second order on an approximately circular neighborhood
and obtains a solution for the best step/peak combination through the
circle.  It is omni-directional. Note that it is primarily sensitive
to the gradient.
 The operator used in the Binford-Horn
edge finder evaluates the laterally-inhibited signal along a variety
of directions.  It is not sensitive to the gradient.
The Hubel¬Wiesel cells are also specialized to specific directions.
[ These cells/operators find edge POINTS, but what kind of organization
could do linefinding/following in parallel? ]

3.  There are known to be stereo cells of the Hubel-Wiesel type
(Blakemore).  These correlate edge signals in the two eyes.

4.  Humans are known to be capable of stereo without features (Julesz).
No cells are known which perform this function; presumably, such exist.
[ By "no features" is meant "no structured, semantic features".  We know
how to program this bulk-correlation.  In fact, Ram has done it!  Arthur
even believes it works!! ]

Conjectured Operators
1.  Corner cells or continuity cells.  These cells check local
smooth connection of edge fragments from Hubel-Wiesel cells.  They
are local in the sense of diameter-limited.
2.  Texture cells

[ DBA: Other points mentioned

1 Suppose you have all these wonderful segmentations available to you in
zero time by magic parallel hardware.  How do you decide where to look
first?  You might think that if you could decide without looking at all
the input data then you could have left the segmentations until later
and done only parts of them.  For example you make a depth map and then
decide to look at the depth at the centre of it.  It would be cheaper
to decide first and then just map the depth at the centre!  So the answer
must be that more global stuff than say Retinex is done by the "parallel"
or "dumb segmentation" part.  For example the region with the largest
value of homogeneity*size*boundary smoothness might be found.

2 What is a curve-detector cell?  Or, for that matter, a line-detector
cell, as opposed to an edge-point detector cell?  What kind of parallel
set-up could do what our serial line-followers do?

3 Arthur said that "brains don't have addresses", which I took to be some
remark about the hardware difference between brains and computers which
he believes to have some bearing on the machines-are-serial-brains-are-
parallel-at-what-level-is-the-program-the-same problem.  DBA did not
understand this remark. ]