00100	define the domain define high and low level
00200	
00300	
00400	Computer visual perception concerns visual systems for function of an
00500	organism such as a robot.   A  typical  problem  which  arises  from  this  task  is:
00600	describe the bodies in a scene, and describe their relations.  The
00700	concern is with  three-dimensional  interpretations,  segmented  into
00800	objects,  with  descriptions of the objects and their relations. 
00805	We want to manipulate or avoid objects, not just classify them.
00850	This
00900	is more ambitious than just  classification  among  a  small  set  of
01000	classes,  which  is the statistical pattern recognition paradigm. The
01100	choice of paradigms is important in the evolution  of  a  discipline;
01200	for  our  purposes, the character recognition paradigm is inadequate,
01300	suggesting the wrong value judgments.
01305	
01310	Why should we be concerned with research in visual perception?
01315	Because of the enormous implications for productivity which the
01320	are possible.  We will describe these in the next paragraph.  Because
01325	of the importance of such problems in AI, learning and computer science in
01330	general.  And because of the significance for the understanding of
01335	biological systems.
     

00100	Research in computer visual perception effectively began 10 years ago
00200	with  Roberts  [].   That  work  set several important directions for
00300	future  research:
00400		the  problem  chosen  was  perception  of   three-
00500			dimensional  objects;
00600		a  complete  system  went  from image input to display
00700			of internal descriptions;
00800		descriptions of objects were built up from volume
00900			models, a block, a  wedge,  and  a  hexagonal  prism.
01000	The system assumed it would first obtain perfect line drawings,  then
01100	interpret  them.  A number of passes transformed line fragments based
01200	on the gradient into chains of fragments, extended in both directions
01250	because it does not require any of the mechanism required for three
01255	dimensional scenes, and because almost any technique will work partially,
01300	by  sequential  rms line fit which segmented at corners. Lines ending
01400	near other lines were joined at vertices.  Model points were  matched
01500	with   topologically   equivalent   image  vertices  by  translation,
01600	rotation, and scaling, followed by a perspective camera transform.
01700	
01800	That system organization has proven inadequate in usual  cases.
01900	In  other  studies,  such  techniques encounter difficulties at every
02000	stage of the sequence: lines for which no fragments are found,  lines
02100	from  shadows  and  reflections,  garbled line drawings from gradient
02200	data that look adequate to the eye, errors compounded at each  pass.
02300	How  can we improve the versatility and reliability of such a system?
02400	It is not enough to improve the individual modules themselves; it  is
02500	not even the most important improvement.  So long as errors compound,
02600	only perfect modules will do.  Instead, a)  use  partial  results  to
02700	guess  new  tests; for example, at vertices, check to see whether one
02800	line continues; or guess missing lines from a  partial  line  drawing
02900	and  verify.  b)  use  high  level  knowledge  and  models to aid the
03000	low-level operation. c) use  a  variety  of  sensory  information  to
03100	determine  consistency  or  contradiction. Use color, stereo, motion,
03200	shading. d)improve the modules. The first three provide alternatives,
03300	and  thus  require  decisions  of consistency or contradiction.  This
03400	changes the pass-oriented  structure  to  one  with  more  connection
03500	between  high  level  and  low  level,  and  which  is  determined by
03600	interpretation of the scene.