Machine vision is "taking off"!
The panelists from industry
(R. Anbalagan of Innovision, Vic Nalwa of Lucent Technologies, and
Tom Olson of Texas Instruments)
were all quite positive about the
future of image computation technology. Not only is it important
in the fields of object recognition and scene understanding, which are
traditional CVPR interests, but it is also important in other areas
such as virtual reality, databases, and networked communications.
From a survey of 60+ persons present in the room, 14 represented
colleges had a course for undergraduates in computer vision or image
processing; 4 had a course in pattern recognition (means that
undergraduates were able to take the class and were anticipated
in the class; some of these courses were for both grads and undergrads).
We take this as anecdotal evidence that there is strong need for
additional educational opportunities for students.
Machine vision is interdisciplinary!
It relates to computer science, mathematics, physics and engineering,
and, most importantly, to various application areas, such as geology,
medicine and manufacturing. While this creates many opportunities in
education, it also makes for difficult choices in designing courses.
Students need more than just theory.
Although there is well-defined computer vision theory --
as some panelists have included in their position statements --
there is a strong need for laboratory work and project work with a
systems approach. Short labs are needed to acquaint students with
cameras, lenses, and lighting. Significant application exposure is needed,
meaning study of a real application problem and development of a systems
approach to its solution. Solutions include sensing and software development
and attention to real-time processing. For example, an engineering student
should consider a line-scan camera as an alternative to a 2D camera for
imaging moving objects.
Where in the curriculum?
The traditional method of introducing undergraduates to
areas that involve image-related computation is through
offering upper-level elective courses. However, many programs
may not have the either (1) the faculty time available to staff
more elective courses, or (2) the available elective hours in
the degree program to allow many students to take the course.
Problems that arise due to constraints of faculty time and
course credit hours in a degree program were not discusses in
this panel, though they may be very real in some cases.
One of the more novel suggestions to receive some favorable response during the panel discussion was that problems involving image-related computation could be used as examples throughout the "core" curriculum. One of the panelists, Dmitry Goldgof, described an example of this in his experience teaching a standard Data Structures course for Computer Science majors at the University of California at Santa Barbara. In this particular course, all of the programming assignments involved image-related computation. (See Goldgof's position statement on the web page mentioned above.)
How to teach
Major areas of agreement about how to teach image-related
computation included the following.
* Use interactive demos that allow visualization of the problem.
One concrete example of this concerned the effect of different
parameters or an edge detection filter. Eric Krotkov
of Carnegie-Mellon University
described a tutorial module which allows the student to view an image,
the shape of a edge detection filter, and the resulting edge detected
image. Slide bars allow the student to interactively change the values
of parameters in the edge detection filter. Through experimenting with
the tutorial module, the student can gain an intuitive grasp of the
effect of different parameters on the final result.
* Give substantial project assignments.
All of the industry participants in the panel
identified this as
a weakness in the backgrounds of the students that they hired.
While all agreed that it is not feasible for students to write
software on the same scale as must be done for real applications
in industry, there was still the feeling that (1)
students could be exposed to larger project implementations than
they typically see in their courses, and
(2) students could be exposed to group-oriented project assignments.
One panel participant pointed out that it is typical for
students in a compiler course to write a compiler for a simple
language, and it is typical for students in an operating systems
course to write or modify portions of a small operating system.
Why, then, could students in a computer vision course not implement
and "signal-to-symbol" system for some idealized problem?
What to teach
Perhaps not surprisingly, the question of what to teach showed
the least clear agreement.
There was agreement that the math/theory side should not
be sacrificed in favor of glitzy demos and high-level packages
that hide the details of how the technique works.
At the same time, the point was made that students cannot
implement even a toy system if they must do it all "from scratch"
in one semester.
There was no clear agreement on software packages, programming
languages, or other such issues.
Areas for future progress
There was general agreement that there are not yet
ideal textbooks for undergraduate electives in most areas
of image-related computation.
The situation for "core" course textbooks which include
examples of image-related computation is even worse --
no one could think of even a single textbook to mention.
The time limitations of the panel discussion prevented
detailed discussion of particular project assignments
and case studies.
At the end of the panel discussion, approximately 30 people
expressed interest in a one-day workshop on educational issues,
to be held in conjunction with a future Computer Vision
and Pattern Recognition conference. Initial planning for
such an event is underway for CVPR '97.