Boyer - 2
to 9 minutes for 512 x 512 images on a Sun SPARC-20 workstation.
This is for edge extraction, segmentation, and subsequent organi
zation of the edge segments into structures such as parallelograms,
rectangles, circles, ellipses, ribbons, and triangles. Each of these
structure hypotheses is associated with a belief measure.
The key notion in gaining this speed is the use of the PIN for top-
down processing in what we call the “attentive phase.” Previous
approaches to perceptual organization have mostly been purely bot
tom up, without any top-down knowledge base influence and there
fore have been entirely dependent on the imperfect input data, and
subject to significant combinatorial thrashing as noted by the Chair
men’s question. The PIN, which acts as a sort of knowledge base,
besides coping with the input imperfections, also allows us to in
tegrate multiple sources of information and to form a composite
organization hypothesis.
Is the perceptual grouping model really useful? Is it actually implemented and
used?
It should come as no surprise at this point that our answers are
YES and YES. We are using it in aerial scene analysis and in other
application domains. In particular, we are now (with Prof. Sudeep
Sarkar, now of the University of South Florida, USA) applying these
methods to the particular problem of change detection in aerial im
agery. Changes in the scene structure are inferred as changes in
the scene’s perceptual organization and can be detected with some
simple statistical tests. That is, the fact that significant structural
changes in the area have occurred can be detected. Analysis of those
changes is, of course, more involved. Changes in the perceptual or
ganization of the scene and, especially, its statistics, are relatively
unaffected by lighting, weather, and viewpoint.
More generally, we argue that the evolution of perceptual organi
zation in biological vision, and its necessity in advanced computer
vision systems, stems from the characteristic that perception is an
intelligent process. This is particularly so for higher order organisms
and, analogically, for more sophisticated computational models. We
believe that one can represent the sophistication and performance
of a vision system (biological or artificial) almost entirely in terms
of its ability to compute perceptual organization.