phyllis@utcsrgv.UUCP (Phyllis Eve Bregman) (02/17/84)
UofT Department of Computer Science Seminar Schedule for
the week of February 20th, 1984
Thursday, February 23th, 3:00 P.M., GB244: Mr. Jan Mulder, Laboratory
for Computational Vision, Department of Computer Science, University
of British Columbia: "Representing ambiguity in visual knowledge".
ABSTRACT: One of the key issues in Computational Vision is the
proper mapping of image features to interpretations. Local image
features are usually ambiguous in this respect. Only by considering
features over a large area can we constrain the set of possible
interpretations for each feature. Most Computational Vision systems
solve this ambiguity problem by a hypothesize-and-test approach.
A different approach is taken in the model-based system described
in this talk.
Multi-rooted discrimination trees are used to represent classes of
objects that can have a similar appearance in the image. At the
leaves of such trees we represent basic objects which describe the
image unambiguously. All other nodes represent classes of such
objects. Rather than invoking basic objects directly, as is done
in the hypothesize-and-test approach, we invoke the class that
represents the complete set of possible objects. Following the
principle of least commitment, we then replace this class by one
of its subclasses as we expand our focus of attention over the image.
This process continues until the image no longer provides information
by means of which we can do further sub-classification.
The process of discrimination can be treated as a constraint
satisfaction problem. A new form of Arc Consistency has been
developed to deal with the hierarchical domain inherent in
discrimination trees.
--
Phyllis Eve Bregman
CSRG, Univ. of Toronto
{decvax,linus,ihnp4,uw-beaver,allegra,utzoo}!utcsrgv!phyllis