Re: Image Vision

John R. Garrett (mailto:0004716758@MCIMAIL.COM)
Fri, 15 Jul 1994 10:36:00 EST

Message-Id: <mailto:199407151654.LAA00874@library.wustl.edu>
Date:         Fri, 15 Jul 1994 10:36:00 EST
From: "John R. Garrett" <mailto:0004716758@MCIMAIL.COM>
Subject:      Re: Image Vision
To: Multiple recipients of list IMAGELIB <mailto:IMAGELIB@ARIZVM1.BITNET>

Paul, one way to think about ordering and retrieving purely visual images is
to look at the experience of the major stock photo agencies, which have been
retrieving based on vague descriptions for many years.  As I understand it,
the photos are organized in files by their main descriptive elements, but
often what is sought doesn't correspond.  So the system depends primarily on
experienced people, who have a powerful visual memory and, when asked for a
photo (e.g.) of a ship in a harbor with a cliff in the background, can find
it in their visual memory, then link that visual memory to a particular
file, although the desired photo may not be stored according to any of the
obvious visual clues.  It might be worthwhile to try to find out how these
folks (who are still at it) actually function, and think about expert
systems.

In this case, humans mediate verbal and visual information. The other path would be to skip the verbal descriptives entirely, and look for purely visual ways of linking need and works. There are a variety of research projects out there (some of which we sponsor) which look to find correspondences among images. But I don't think (by definition, I suppose) we know how to (or perhaps ever can) even talk about this problem. I am convinced, however, that just layering on more descriptive words (as in the California photo archives, which I greatly admire) just won't scale in the long run, and avoids the real problem.

That said, I have no clear idea how to say what the real problem is.

John Garrett