Basic Research into Vision

A.  Goals stated in the proposal for 1976-77

	1. Unify implementation of low-level visual processing and aggregation code.

	2. Develop a theory of lightness computation from the primal sketch, and
	   implement it. This is a prerequisite for analysis of color.

	3. Test the texture discrimination theory with experiments
	   on appropriately selected micro-texture arrays.

	4. Arrange to handle larger images, up to 256 by 256 pixels.

	5. Develop good algorithms for very local organization in the
	   image (blobs and thin lines).

	6. Develop and implement a theory for extracting axes from
	   an image prior to its three-dimensional interpretation.

All of these have either been completed, or are proceeding in a satisfactory manner.


B. Additional theories and discoveries

	1. Representation of three-dimensional shape.
	   New theory based on the idea of decomposing a shape into connected
	   generalized cones.
	   This representation is simple, powerful, requires little computing
	   power to maintain it, and is connected to the axis-theory mentioned
	   above.
	   Furthermore, we have evidence that this is how humans represent shape.

	2. Cooperative computation of stereo disparity.
	   (a) Solves a long outstanding computational problem.
	   (b) Solution valid for a very general class of images
	   (c) A prerequisite for automatic stereointerpretation of aerial photos.
	   (d) First use of a radically new style of computing, which we
	       have named cooperative algorithms.
	   (e) Likely to have wide application to other problems in early visual
	       processing, like the discovery of texture flow patterns in an
	       image, the correspondence between successive frames taken with
	       a TV from a moving vehicle.

	3. Structure from motion theory.
	   (a) Again, solves a long outstanding computational problem.
	   (b) Solution valid for rigid bodies, i.e. for a very general class
	       of images
	   (c) Problem solved is the calculation of an object's shape from
	       information about the way its image changes as the object rotates
	       relative to the viewer.

C. General Observations

	This work is central and vital to an overall theory of visual information
	processing. One cannot hope to have a general-purpose vision system until
	we have laid a solid theoretical foundation for what needs computing,
	and how. This foundation is emerging rapidly. The subject is in an extremely
	promising and fruitful period, and we expect rapid progress to continue.ββ