Information Technology Reference
In-Depth Information
Want to read more?
This chapter was about interpreting visual scenes; it showed how thinking played a
fundamental role in that process. (The picture in Figure 6.2 is taken from [35].) The
aerial sketch map interpretation of section 6.2 was inspired by the work of Reiter
and Mackworth [34], and the object recognition task of section 6.4 has appeared
in many guises and is the subject of a textbook by Grimson [32]. Other studies of
knowledge-based vision include the early work of Badler [29] and Tsotsos [36], con-
cerning motion in visual scenes. Another related area is reasoning about space and
how shapes occupy it (see, for example, [30]). For a retrospective on the problem of
visual interpretation as a whole, including some of the challenges faced by researchers
in the field, see [31]. From a more artistic point of view, the idea of impossible objects
and visual ambiguity is exploited to great effect in the work of M. C. Escher (see, for
example, his
Relativity
).
However, much of what happens when we view the world around us appears to
take place without any knowledge of the world, and therefore (by the account here)
without any thinking. For example, we do not have to know what we are looking
at to spot something yellow in a large field of red. Pylyshyn [33] has written a book
explaining how this part of the visual process is very different from thinking. The
subarea of AI called
computer vision
is currently dominated by research on “early
vision” of this sort.
The polyhedral scene labeling considered in section 6.3 is an interesting intermedi-
ate case. It comes out of the work of Waltz [37] in the 1970s and was one of the first
tasks formulated as a constraint satisfaction problem. However, it is not necessary to
know very much about the world to perform this task. There are vertex types and line
labels but no open-ended class of objects like vehicles or houses or banquet tables. For
this reason, Pylyshyn prefers to categorize this task as not depending on thinking at
all, but rather as embodying what he calls “natural constraints.”
If you reflect on what it actually feels like to look at visual scenes, you might be
tempted to conclude that the thinking is really quite unlike constraint satisfaction. It
seems like visual interpretations come to you without any mental effort—nothing at
all like solving a Sudoku puzzle, for instance. Part of this feeling is due to the fact that
your eyes are able to quickly dart back and forth over a scene, too quickly to register
how you are using what you know. To get a better sense of how a visual interpretation
is assembled, imagine that you cannot move your eyes in this way. Simulate this by
cutting a small hole in a sheet of paper, looking at an image like figure 6.2 through
the hole only, and then slowly moving the paper to take in the entire scene.
