Mentoring Undergraduates In Computer Vision Research
Mubarak A. Shah and Kevin W. Bowyer
Abstract
The future of our society will be shaped by the young and
talented minds going through our colleges and universities today. During the
last ten years we have worked with over seventy undergraduate students from a
half dozen institutions in Florida under Research Experiences for Undergraduates
(REU) program funded by NSF. A large fraction of our REU participants have been
able to prepare a paper for submission to a conference, have the paper accepted
and then attend the conference to present the paper. Several participants have
even accomplished substantial enough research to also result in journal
publications. Many of our past participants are now pursuing graduate studies at
various institutions. In this paper, we describe our REU model in detail,
discuss some examples of student success, and summarize our observations.
Introduction
The NSF predicts a shortfall of around 700,000 scientists
and engineers by 2010. In 1977, there were 4,000,000 high school sophomores in
the U. S., 730,000 of these students expressed interest in science and
engineering careers. In 1980, when this cohort entered college, 340,000 retained
their interest. By 1984, 206,000 had actually graduated in scientific or
engineering disciplines. Only 61,000 of these men and women entered graduate
school in science or engineering. By 1992, just 9,700 will graduate with Ph.D.s.
To have only 0.2% of these students end up with doctoral degrees does not seem
promising! On the other hand, according to the report by the U.S. department of
labor, between 1986 to 2000 the computer and data processing industry is
projected to be the fastest (76%) growing of all industries in the economy. The
Chronicle of Higher Education for February 26, 1997 states that: ``The United
States may lose its lead in the information- technology industry because it
isn't producing enough educated people to work in the fast-growing field, a
report released recently warns.''In 1987, NSF
started a new program, Research Experience for Undergraduates (REU).
The goal of the REU program is to encourage talented students to pursue graduate
studies and realize their full potential in this regard. We have received six
NSF REU site awards totaling over $800,000 during the last ten years. The first
grant was in 1987, with the first author as the sole P.I.. Since then we have
received grants in: 1989 (P.I. Mubarak Shah, Kevin Bowyer, G. Krishnan), 1990 (P.I.
Mubarak Shah, Kevin Bowyer, G. Krishnan), 1991 (P.I. Mubarak Shah, Kevin Bowyer,
G. Krishnan ), 1992 ( P.I. Mubarak Shah and Kevin Bowyer) funded for two years,
and 1995 (P.I Mubarak Shah, Kevin Bowyer, Louise Stark and Niels Lobo) funded
for three years. Approximately one hundred undergraduate students from a half
dozen institutions in Florida (e.g., Rollins College, Stetson University, Eckerd
College, UCF, USF, FIT) have participated in this program. This paper describes
our experience with NSF-funded REU projects over the last ten years.
General Philosophy
We believe that the best way to achieve the goal of the
REU program is to involve the students as a part of a successful, active
research group. This exposes them to the intellectual excitement that is
involved in research, encourages them to think creatively and independently, and
helps them to develop the skills necessary to work on research projects. In
addition, students are exposed to professional meetings, learn to assimilate the
latest research from reading and discussing recent papers, and learn how to
write up and present the results of their own research.
It is obvious that an undergraduate research experience
that lasts only a few months does not allow sufficient time to complete a
serious project. Hence, such a short project would not allow participants the
chance to achieve the feeling that they have made a significant original
contribution. Since this feeling of accomplishment is the primary motivation of
most researchers, we believe that the REU experience must be structured so that
each student has the maximum opportunity to successfully complete their chosen
project. Hence, our REU experience is spread out over (at least) one calendar
year, typically as a summer/fall/spring sequence.
From our experience, we have found that interaction
among students from different institutions greatly increases the diversity of
ideas and the quality of the experience. During the academic year, however,
guidance and supervision is difficult if the student and the faculty member are
at different institutions. Consequently, we feel that it is important to have a
faculty member at each institution directly involved with, and responsible for,
each student. The faculty member can then easily supervise the student's
academic year course load and be available for advice and guidance. We feel that
our management strategy for the program has been demonstrated over the past ten
years to be quite successful.
Selection of Participants
We primarily look for students who are juniors or seniors
in the Fall semester of each year, who have a GPA of at least 3.25, and who have
a strong background in Mathematics and basic Computer Science. After an initial
screening based on written applications, the faculty members meet individually
with the potential participants and decide which should receive offers.
While we have found that ``word of mouth'' from current
and past participants is the most effective recruiting tool, the following
efforts are also made:
- advertisement of the opportunity in the campus
newspapers,
- notices posted at various places on the campuses,
- announcements made in the junior level classes,
- announcements distributed in Tau Beta Pi
(Engineering honor society), UPE (Computer Science honor society), IEEE and
ACM student chapters,
- juniors majoring with a GPA of 3.25 are sent a
letter inviting them to apply,
- students in the {\em Honors} programs are
contacted, and encouraged to apply,
- current REU participants are asked to give a talk
about their experience at SWE (Society of Women Engineers) and SME (Society
of Minority Engineers) meetings in the spring semester, Faculty/Student
Seminars, and
- we personally contact women, minority and
especially-talented students and encourage them to apply.
On the average, every year we receive about twice as many
applications as the number of students we can select to participate. However,
that number has recently been decreasing because good employment opportunities
exist for computer science majors. The students who are not selected to
participate typically either (1) do not have enough background in Mathematics
and programming, or (2) are unable for academic and/or financial reasons to
devote subsantial time in the summer to the REU program.
Student Activities
One of the important goals of this experience is sparking
the student's interest in scientific research as early as possible through
student-faculty and student-student interaction. At the same time, we feel that
each student must have sufficient time to mull over problems and somewhat
independently achieve a reasonable level of solution of a problem of their own
choosing. Students are able to devote a major portion of their time to the
project during the summer, since they are not involved in regular courses.
During the academic year, continued student involvement is ensured by the
faculty member at each institution. The professor(s) at each institution
supervise the students through their registration in elective Computer Vision,
independent study and senior project courses and through weekly individual and
research group meetings.
Our last proposal to NSF for REU covers a period of
three years. Our experience has shown that there is a need for flexibility to
deviate from the ``standard'' one calendar year in certain situations. Sometimes
an excellent prospective participant has an academic or co-op schedule which
does not allow a continuous 12-month participation. In such cases, it would be
useful if the ``REU year'' could be spread over a longer calendar period.
Sometimes, also, one of our participants does excellent work on his/her topic
and does not graduate for as much as another year after the end of the 12-month
REU schedule. In these cases, it is useful to offer the student a chance to
stretch their ``REU year'' to some longer period. Lastly, a participant may be
forced, due to personal reasons, to drop out of the program, and then a new
participant must be found to take their place. We regularly see perhaps three
special cases out of ten participants.
Student-Faculty Interaction
The appropriate level of student-faculty interaction and
supervision of the participants is achieved by the following means:
- Weekly Individual Meetings: Each faculty
holds a weekly individual meeting with each of the REU participants to
discuss their research project.
- Weekly Research Group Meetings: Each
faculty holds a weekly seminar-style meeting with his research group, in
which current research papers are discussed, upcoming presentations by group
members are rehearsed and discussed, papers by group members are
``pre-reviewed," participants give presentations on the status of their
project, etc.
- REU Group Meetings: The whole REU group,
consisting of faculty and REU students meet once a semester. This meeting
lasts for one day and is scheduled on weekends. The meeting is structured
into three sessions consisting of presentation, demonstration, and
discussion sessions. During the first session the student present their
projects. These presentations include a brief introduction of the problem,
results to date, and future plans. Next, the students are able to
demonstrate their projects in the Computer Vision lab. Finally, we have an
open discussion session where each student is encouraged to comment on other
projects.
- Electronic Mail: Students are able to
communicate with fellow students, the graduate research assistant and
faculty at their campuses as well as at other campuses using electronic
mail.
Journal Club
We believe that REU participants should also be exposed to
other research problems in Computer Vision, beyond their own area of
concentration. In order to achieve this, we organize a ``journal club'', a paper
discussion session during the spring semester. By the start of the spring
semester, students have a fairly good idea of basic concepts in computer vision.
Each time a research paper is selected and assigned to a student. That student
study the paper carefully and present it in the journal club. The other students
also read the paper and prepare questions and comments for discussion with the
student responsible for that meeting. That way students will follow the content
in the paper, and will also be trained to be critical in a constructive sense.
These sessions are moderated by the faculty at their respective universities.
This is an important and effective means of getting students used to making
presentations about technical material and of teaching them how to read and
comprehend journal papers.
REU Speaker Series
The aim of this series is to allow the REU participants to
meet and interact with some well-known researchers in the field. Typically,
during the morning the speaker gives a talk on some aspect of his or her
research. We arrange a lunch or dinner for the speaker to which the whole group
of REU participants are invited. Finally, in the afternoon we hold an informal
discussion session where the REU participants are free to ask questions related
to any topic in computer vision. We attempt to schedule at least two guest
speakers every year.
Participation in Professional Meetings
Attendance at professional meetings gives the REU
participants exposure to the well-known researchers in the field, provides an
opportunity to see good and not-so-good research presentations, and provides an
opportunity to assimilate the latest research results. Therefore, we take REU
participants to at least one professional meeting related to computer vision. We
are fortunate in that our geographic location make this possible at a reasonable
cost. The annual SPIE Intelligent Information Systems meeting held in Orlando
each spring has at least one conference devoted to robotics and computer vision.
This is our default option. In many years, a more focused meeting is held close
by. For instance, in 1994 and 1996 the IEEE Workshop on Applications of Computer
Vision was held in Sarasota, and in 1995 the IEEE Workshop on Computer Vision
was held in Miami. Sometimes the students are also taken to a conference outside
Florida, for example REU students were taken to ICCV9 '95 in Boston. REU
students who get a paper accepted in a conference are often sent to the
conference with travel funds from the grant.
Student Preparation and Follow-Through
Students receive background preparation for their research
project through the computer vision short course in the summer. All students
devote a substantial portion of their time during the academic year to the
research project, since they have an independent study, senior project or senior
thesis supervised by one of the faculty. We help the students as they apply for
graduate school admission during the academic year. We encourage them to take
the GREs early enough that they can take them again if they do not do well
initially. (It has been our experience that students taking the exam a second
time often improve their score by over a hundred points.) We advise them about
which schools have good programs in what subareas. We also encourage qualified
students to apply for graduate fellowships awarded by NSF, NASA, ONR, AFOSR, the
Florida Endowment Fund program, and the State of Florida. We facilitate
interaction with outside researchers visiting any of the campuses during the
academic year, so that students develop the contacts necessary for admission to
the best schools. We also contact company representatives and have them arrange
interviews and site visits for students interested in industry.
Yearly Schedule
June 1 -- July 15
The first major activity for the students is the short
course in Computer Vision. Since, we assume that the students do not have prior
background in Computer Vision, this course quickly introduces them the subject.
The topics covered include: imaging geometry, edge detection, region
segmentation, 2-D shape, stereo and shape from shading, motion, etc. The lecture
component of this course is a one-day meeting each week for six weeks. These
lectures are shared by Professors Shah and Bowyer. During the remainder of each
week, the students re assigned some readings in the area covered by the lecture
material and be given one or two short programming projects which support the
material. In the first two weeks, some extra instruction in C programming, X
windows and use of the Unix workstation are given. This method of scheduling and
presenting the short course was arrived at after review of the 1989/90 project
and discussions with the participants about the effectiveness of the six
straight days format versus the one day a week for six weeks format.
July 15 -- August 15
In the third week of July, each of the faculty members
suggest a number of possible projects and research topics. Each student is able,
within some limits, to {\bf choose} a project topic rather than have one {\bf
assigned}\/. We feel that this is important in terms of having the student feel
that the project is ``theirs.'' Also, we are careful to structure projects so
that they have several possible levels of completion. This ensures that each
student will be able to complete some (possibly simplified) version of their
project and experience the resulting feeling of accomplishment. It also allows
the more talented or ambitious student who works hard enough to achieve a
substantial research publication. This approach has proven successful with our
past REU participants. After each student has a specific project, student is
then given background reading assignments and program projects for their chosen
project. By the end of this period, they should be prepared to write a detailed
prospectus of their proposed research project.
August 15-- August 30
In this time period, we hold joint meetings at which each
REU participant present their proposed research project plan. The other members
of the group provide comments and constructive criticism. Then the participants
are able to revise their plan and pursue their projects full-time until the
beginning of the fall semester.
Academic YearDuring the Fall and
Spring semesters, the students continue working on their individual projects.
There is one meeting of the entire REU group during each semester, to review the
projects and discuss the problems encountered and the progress made. The
students make presentations of their projects, and when appropriate demonstrate
their results on workstations. Further, the participants be taken to at least
one conference at which Computer Vision is a major topic. At the end of the
year, each student is asked to write a comprehensive report describing the
project. Students with publishable results are encouraged to write conference
and/or journal papers.
Student Success
Approximately one hundred undergraduate students from a
half dozen institutions in Florida (e.g., Rollins College, Stetson University,
Eckerd College, UCF, USF, FIT) have participated in this program. Many of these
students have co-authored papers, and most of them have been admitted to
graduate programs. Here are some examples.
- A 1994/95 REU participant, Shawn Dettmer,
worked on Lipreading project. His paper was accepted in {\em
International Workshop on Automatic Face and Gesture Recognition}, Zurich,
1995. He was admitted to the graduate School at UCF. He received a DSR
enhancement fellowship for two years. He completed his M.S. in spring of
1997. Currently, he is a Ph.D. student.
- A 1993/94 REU participant, Jim Davis,
worked on gesture recognition for his REU project. A short version
of the paper about Jim's work was accepted in 1994 European Conference
on Computer Vision (acceptance rate 18%), and a long version appeared
in a journal. He completed his Honors in The Major thesis in 1994.
Jim was admitted to the graduate program at MIT with full fellowship.
He completed his M.S. in 1996. Currently, he is a Ph.D. student at MIT Media
Lab.
- A 1993/94 REU participant from USF, Chris Doss,
worked on ``Comparative evaluation of pattern recognition techniques for
detection of microcalcifications in mammography''. His paper appeared in
International Journal of Pattern Recognition and Artificial Intelligence.
Christopher is now in the graduate program at North Carolina State
University.
- Another 1993/94 student Warren Macchi worked
on ``Interreflections with rough surfaces'', and his paper appeared in 1994
International Conference on Pattern Recognition. Warren received a three
year fellowship from NSF for his graduate degree at UCF. He passed
his Ph.D. qualifying exam in the Spring of 1997.
- A 1992/93 REU participant from Eckerd College,
Chris Stimac, worked on ``Refinements of DCN neural network
algorithm.'' Chris entered graduate school in Mathematics at UC Berkeley.
- A 1991/92 REU student from USF, Melanie
Sutton, worked on ``Reasoning about function to achieve generic
recognition of rigid 3-D shapes.'' This led to graduate work which was
eventually published in the journal Pattern Recognition, and received
honorable mention for Pattern Recognition's best paper of 1994 award.
Melanie is now on the faculty of Computer Science and the University of West
Florida.
- A 1989/90 REU participant, Matt Lavoie, who
received the Outstanding Research award for his Honors in The
Major thesis (REU project), was invited to make a presentation on his
REU project at NSF in March 1990. Matt published one journal and one
conference paper. He successfully completed his MS at UNC, and now co-owns a
company.
- A 1989/90 REU participant, a USF student, Maha
Sallam, had a paper describing her REU work accepted by the Pattern
Recognition Letters journal and then had a paper describing her
continued work accepted as a ``long paper'' at the 1990 International
Conference on Computer Vision (ICCV) (acceptance rate 5%). She
completed her Ph.D. at USF and is working on a medical imaging start-up
company in the Tampa Bay area.
Observations
Here are some observation from ten years of experience
with REU.
- The department and University strongly encourage
this activity. Each year we receive some matching funds from the University
for this grant in terms of course release during the academic year, graduate
student support, travel support etc. The University newspaper has published
at least four stories about REU, twice on the front page. In 1991, we had
had a luncheon for honoring REU students with president and dean in
attendance. REU is considered positively for the tenure and promotion
process, and some other awards (e.g., Teaching Incentive Award (TIP)). REU
stands out from other standard things faculty do: teaching, research and
service.
- From the faculty point of view, REU is very time
consuming. In particular, summer is very hectic. We are tied up to stay
almost whole summer at our campus, because it is crucial for REU students.
The problem is that the REU grant does not pay for any part of summer
salary; we have to get support from other grants, or teach during the
summer. The direct benefit of this is that on the average every year at
least one student produces a good paper. We also get potential graduate
students.
- Some students are either discontinued or drop out
from the program. Some of those students find out that they are not really
interested in research, and others find that financial concerns are playin a
big role in their life and that they can make more money at a part-time job.
- We have realized that there are two main
deficiencies in the students' background relative to preparation for
research: Mathematics and Programming.
- We feel summer is the crucial time, because during
fall and spring students get very busy in their course work. If a student
does well during the summer, and have defined a good project, and started to
get some results, then he or she will do well during fall and spring.
- We have observed that some bright students are
only interested in superficial things. They will write some code to
experiment something. But, when things get serious, and become mathematcial
they back off.
- Some students do not do that well during REU, but
later they realize their potential, and become heavily interested in
research, and go to graduate school.
List of Journal Papers and Book Chapters by REU
Students
The following is a list of publications co-authored by
past participants in the REU program. The names of undergraduates are shown in
bold face.
- James Cryer, Ping-Sing Tsai and Mubarak
Shah, Combining Stereo and Shading, Pattern Recognition, Volume 28,
No. 7, pp 1033-1043, July 1995.
- Li Nan, Shawn Dettmer, and Mubarak Shah,
Visually Recognizing Speech Using Eigensequences, chapter in
Motion-Based Recognition Kluwer Academic Publishers, 1997.
- Woods, K.S., Doss, C.C, Bowyer, K.W., Solka,
J.L., Priebe, C.E. and Kegelmeyer, W.P., Comparative evaluation of pattern
recognition techniques for detection of microcalcifications in mammography,
International Journal of Pattern Recognition and Artificial Intelligence,
pp 1417-1436, December 1993.
- J. Davis and M. Shah, Visual Gesture
Recognition, IEE Vision, Image and Signal Processing,
141(2):101--106, April 1994.
- P. Tsai, M. Shah, K. Keiter, and K.
Kasparis., Cyclic Motion Detection. Pattern Recognition , Vol. 27,
No. 12, 1994.
- Jay Hackett and Mubarak Shah, ``Multisensor
Fusion: A perspective'', chapter in Trends in Optical Engineering,
1993.
- J. Hackett, M. Lavoie, and M. Shah. Object
recognition using multiple sensors. Journal of Information Sciences and
Technology, 1992.
- Krishnan Rangarajan, Bill Allen, and
Mubarak Shah, ``Matching Motion Trajectories'', Pattern Recognition,
Vol. 26, No. 4, pp 595-610, April 1993.
- J. Hackett, M. Lavoie, and M. Shah, ``Three
Dimensional Object recognition using multiple sensors'', Journal of
Information Science and Technology , October 1992.
- J. Hackett, and M. Shah. Segmentation using
range and intensity data. Optical Engineering, pages 667--674, June
1989.
- K. Gould, K. Rangarajan, and M. Shah,
``Detection and representation of events in motion trajectories'', chapter
in Advances in Image Processing and Analysis, (editors: Gonzalez
and Mahdavieh), Optical Engineering Press, June 1992.
- M.Y. Sallam, and K.W. Bowyer. Generalizing
the Aspect Graph Concept to Include Articulated Assemblies, Pattern
Recognition Letters, pp 171-176, March 1991.
- Sutton, Melanie, Stark, L. and Bowyer, K.
Reasoning about function to achieve generic recognition of rigid 3-D shapes,
accepted to appear in Pattern Recognition.
- Bowyer, K.W., Sallam, Maha, Eggert, D. and
Stewman, J.H. Computing the generalized aspect graph for objects with moving
parts, IEEE Transactions on Pattern Analysis and Machine Intelligence,
pp 605-610, (June 1993).
- Woods, K.S., Doss, Chris, Bowyer, K.W.,
Solka, J.L., Priebe, C.E. and Kegelmeyer, W.P. Comparitive evaluation of
pattern recognition techniques for detection of microcalcifications in
mammography, International Journal of Pattern Recognition and Artificial
Intelligence.
Mubarak Shah
Last modified: Wed Oct 1 09:31:51 EDT 1997