00100	COMMENT ⊗   VALID 00002 PAGES
00200	C REC  PAGE   DESCRIPTION
00300	C00001 00001
00400	C00002 00002	TOB 10 Oct
00500	C00004 ENDMK
00600	C⊗;
     

00100	TOB 10 Oct
00200	1. In a biological system what are good places to start using
00300		context?  Computer system?
00400	2. What can be parallel and what can be serial?
00500	3. What has been accomplished?
00600	4. What will be accomplished?
00700		5 years
00800		10 years
00900	5. Semantics: specific visual examples?
01000	6. How to do x?
01100	7. What is a diagram of a good vision system within the
01200		state of the art?
01300	8. What incremental improvements are possible for current systems?
01400	9. Is the notion low-level defined?
01500	10.Is the notion high-level defined?
01600	DBA 18 Oct
01700	These are papers we could discuss
01800	11. Arthur's thesis proposal
01900	12. Marty Tenenbaum's recent SRI memo
02000	13. Feldman/Yakimovsky on decision theory in vision
02100	14. Max Clowes' notes for his book
02200	15. Recent MIT vision stuff - Minsky's frames, Freuder's advice
02300	16. Recent MIT vision stuff - Marr and Horn
02400	
     

00100	Is The Low-Level defined?
00200	
00300	
00400	There has been a strong emphasis in AI on knowledge and using knowledge
00500	for finding features such as edges.  Is the low-level even defined?
00600	Two questions:
00700		Are there useful features which can be calculated in a context-
00800		free way?
00900		Is there a parallel way of calculation?
01000	The revolutionary progress of semiconductor hardware raises significant
01100	questions; we cannot now say what would be useful on the basis of
01200	simulation of parallel algorithms.  Many of our considerations have
01300	been based on "current machines are almost fast enough.  What we need
01400	are clever programs which do not do much low-level work".  
01405	There has clearly been an emphasis on optimal decision structures which
01410	do no unnecessary work.  Optimality has a very different sense when
01415	parallel hardware is cheap, and for systems with many goals and a
01420	complex universe.  There is not a good analogy we feel, with current
01425	examples of goal-oriented systems which have very limited goals
01430	and limited abilities, simple universe.
01900	
02000	We will attempt to answer the question by examples of systems which
02100	do appear to be low-level in computer vision and biological vision.
02200	
02300	1.  Lateral inhibition--Retinex
02400	A word of explanation.  This system eliminates the effects of smooth
02500	gradients of lighting (including spectral variation) for color vision
02600	and edge detection operation.  Three phenomena are relevant here:
02700	color constancy, brightness constancy, and the Cornsweet effect.
02800	The last is strong experimental evidence that the human visual system
02900	does not find edges at large gradient of intensity, but only discontinuity
03000	in intensity or its gradient.  The teleological advantage is clear:
03100	reflections and curved surfaces are common in nature, and using a gradient
03200	operator as is usual in computer vision would give us whole surfaces of
03300	spurious edge points.  Lateral inhibition has been used in the edge finder
03400	used in the MIT copy demo (Binford), and the papers by Horn and Marr
03500	are recommended.
03600	
03700	Since the system appears to be implemented in the retina, and appears
03800	not to be sensitive to gradients, there is a question of how people
03900	see gradients in surface apparent brightness, which are used for
04000	guessing curvature.  
04100		There may be a non-retinex channel, particularly in the fovea
04200		which may have some one-one connections to the ganglion.
04300		They may be inferrable from the Retinex process.  Consider a
04400		non-uniform surface on a uniform surface.  When we perform the
04500		Retinex operation, the contrasts at the various portions of the
04600		boundary are different.  When the inverse computation is performed
04700		the resulting signal is non-uniform, because of the variation
04800		along the boundary.  Thus, a non-uniform reflectivity does show
04900		up while a non-uniform lighting does not show up, just because
05000		the reflectivity change manifests itself at discontinuities,
05100		while lighting changes don't change values along discontinuities.
05200	
05300	
05400	2.  Hubel-Wiesel cells----edge operators
05500	There is good reason to think that the Hubel-Wiesel cells are edge
05600	operators; there are questions about the other assumptions of their
05700	study, that these cells are innate and that there is a hierarchy of simple,
05800	complex and hypercomplex.
05900	There are several examples of such edge operators: the Hueckel operator
06000	evaluates terms through second order on an approximately circular neighborhood
06100	and obtains a solution for the best step/peak combination through the
06200	circle.  It is omni-directional. Note that it is primarily sensitive
06300	to the gradient.
06400	 The operator used in the Binford-Horn
06500	edge finder evaluates the laterally-inhibited signal along a variety
06600	of directions.  It is not sensitive to the gradient.
06700	The Hubel¬Wiesel cells are also specialized to specific directions.
06800	
06900	3.  There are known to be stereo cells of the Hubel-Wiesel type
07000	(Blakemore).  These correlate edge signals in the two eyes.
07100	4.  Humans are known to be capable of stereo without features (Julesz).
07200	No cells are known which perform this function; presumably, such exist.
07300	
07400	
07500	
07600	Conjectured Operators
07700	1.  Corner cells or continuity cells.  These cells check local
07800	smooth connection of edge fragments from Hubel-Wiesel cells.  They
07900	are local in the sense of diameter-limited.
08000	2.  Texture cells