COMMENT ⊗   VALID 00003 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	.xgp
C00008 00003	ββ
C00009 ENDMK
C⊗;
.xgp
.FONT 0 25VG
.FONT 1 25VRI
.FONT 2 37vrb
.sp 3
.center
ε2COMPUTING LIGHTNESS FROM THE PRIMAL SKETCH
.sp 6
.center
by
.sp 3
.center
D. Marr and R. Woodhamε0
.sp 8
.he1
Marr & Woodham
.he2
LIGHTNESS FROM THE PRIMAL SKETCH
.double
.center
The Artificial Intelligence Laboratory,
.br
.center
Massachusetts Institute of Technology,
.br
.center
545, Technology Square,
.br
.center
Cambridge, Mass. 02139, U.S.A.
.sp 4
	Running title: lightness from the primal sketch
.pa
.single
.center
ε1Notesε0
1. Introduction.
  (a) why another algorithm?
	dissatisfaction with Horn's -- errors wwth difference signal,
		plus the half-sided Cornsweet, plus the expense
	difficult to reconcile (especially colour theory) with the reported
	facts about retinal receptive-field properties. If possible, we
	want a theory based more on Barlow's suggestion, where somehow
	its the relative color contrasts that give rise to the cooputation.
	This stands against Land's method in that the three colors are
	mixed early rather than late -- trying for a 2-D (color triangle)
	version o fthe one-D lightness idea

2. At what stage is perceived brightness computed?
	(a) post-retinal -- because of very high-level effects- 
	   (get that ref from Whitman Richards)
	(b) At least spme part of it is pre-regions, (the H**
	    illusion and so on), but things like gloss etc. suggest
	    that even if the computatin is largely complete earlier than this,
	    there do exist region-based effects as well.
	(c) Question naturally arises, is it possible to compute
	   lightness from the primal sketch in some way?

	   n.b. advantage - that yu can do the bulk of it before breaking
	   up into regions, and can therefore use lightness (rather than
	   contrast) to help in this process. So if it is possible, it
	   would certainly be interesting.

3. Computing lightness from the primal sketch.
	(a) Take out small things - blobs and bars
		-- perhaps mention perceptual evidence here
	(b) Left with essentially one-D info in a 2-D problem, therefore
	   the method has to be 1-D . This implies
	   that it involves a scan of some kind  -- i.e. perp to the
	   items in the sketch.
	(c) Boundary conditions and the Closed Orbiit theorem.
	(d) non-integrating scan algorithm, add orbit thm to get
	   the integrating orbit theorem.

4. Results and examples.
	(a) show several images in which the comp has been done
	(b) Show a Cornsweet edge
	(c) Grid illusion
	(d) do an image in which lightnss depends a bit on an
	   "illusory" contour, like the "sun" effect.

5. Illusions related to brightness
	(a) describe the illusions
	(b) Criticise the centre-surround r/f explanation
	(c) Say what our method might avcount for, and what i does not.
	(4) mention the Fourier lines - perhaps run one
	   of those exploded chess patterns through with large mask sizes
	   to see what happens
	(e) Conclude tht theye are probably several causes for the illusions

6. Discussion
	(a) Analoguish thing, not precisely bound to regions, only roughly
	(b) Effects of fixed scan lines - mention migraines?
	(c) Blobs and bars
	(d) NPHYS tests - orbits& stripes; also when in primal sketch
	   is this done - try to get that info on timing mentioned by
	   Kasday(??)
	(e) Extension to thery of color computation

.end

ββ