Preprint Available

[Bartlett Mel]

Announcing a new preprint, available at:

	url=ftp://quake.usc.edu/pub/mel/papers/mel.seemore.TR96.ps.gz
		(22 pages, 1.1M compressed, 34M uncompressed)

Sorry, no hardcopies.  
Problems downloading/printing?  Please notify author at mel at quake.usc.edu.

		 --------------------------------------

	SEEMORE: Combining Color, Shape, and Texture Histogramming in
	  a Neurally-Inspired Approach to Visual Object Recognition

			       Bartlett W. Mel
		      Department of Biomedical Engineering
		   University of Southern California, MC 1451
			  Los Angeles, California 90089
			
				    ABSTRACT

	Severe architectural and timing constraints within the primate visual
	system support the hypothesis that the early phase of object
	recognition in the brain is based on a feedforward feature-extraction
	hierarchy.  A neurally-inspired feature-space model was developed,
	called SEEMORE, to explore the representational tradeoffs that
	arise when a feedforward neural architecture is faced with a difficult
	3-D object recognition problem.  SEEMORE is based on 102 feature
	channels that emphasize localized, quasi-viewpoint-invariant nonlinear
	receptive-field-style filters, and which are as a group sensitive to
	multiple visual cues (contour, texture, and color).  SEEMORE's
	visual world consists of 100 objects of many different types,
	including rigid (shovel), non-rigid (telephone cord), and statistical
	(maple leaf cluster) objects, and photographs of complex scenes.
	Objects were individually-presented in color video images under stable
	lighting conditions.  Based on 12-36 training views, SEEMORE was
	required to recognize test views of objects that could vary in
	position, orientation in the image plane and in depth, and scale
	(factor of 2); for non-rigid objects, recognition was also tested
	under gross shape deformations.  Correct classification performance on
	a testset consisting of 600 novel object views was 97% (chance was
	1%), and was comparable for the subset of 15 non-rigid objects.
	Performance was also measured under a variety of image degradation
	conditions, including partial occlusion, limited clutter, color-shift,
	and additive noise.  Generalization behavior and classification errors
	illustrate the emergence of several striking natural shape catagories
	that are not explicitly encoded in the dimensions of the feature space.