According to our analysis of Face Recognition research
over the past decade, the scientist whose work ranks at #2
by total citations and #1 by cites per paper is Dr. P.
Jonathon Phillips, with 25 papers cited a total of 890
times. In
Essential Science IndicatorsSMfrom
Thomson
Reuters, his work can be found in the fields of
Computer Science and Engineering.
Dr. Phillips hails from the National Institute of Standards and
Technology (NIST), where he is Program Manager for Face Recognition. He is
also a fellow of the International Association of Pattern Recognition.
In the interview below, he talks
with ScienceWatch.com about his career in face
recognition technology.
Would you tell us a bit about your
educational background and research experiences?
I have a BS in Mathematics, an MS in Electronic and Computer Engineering, a
Ph.D. in Operations Research, and my Ph.D. adviser was a prominent
statistician. The interdisciplinary nature of my education gave me an
appreciation and understanding of fields closely related to face
recognition. In my job, I am simultaneously a researcher and a government
program manager. The program manager aspect of my job makes me a better
scientist by keeping me focused on the big picture.
How did you get involved in facial recognition
research?
"Many proposed face recognition
applications will either replace or assist
security guards."
In the fall of 1992 I was working on my Ph.D. and employed by the US Army
Research Laboratory (ARL), and I had an arrangement with my managers that
my dissertation could be part of my job. In January 1993, the ARL started
to manage the FacE Recognition Technology (FERET) program for DARPA. I
jumped at the opportunity to be program manager for the FERET program and
to conduct research on face recognition for my dissertation. Managing the
FERET program was a great experience that allowed me to get involved in
automatic face recognition just as the field was taking off.
Your most-cited paper in our analysis is the 2000
IEEE T Pattern Analysis paper, "The FERET evaluation
methodology for face-recognition algorithms," which has been cited
close to 400 times. Would you walk our readers through this paper and
why you think it is so highly cited?
The primary goals of the FERET program were to determine if automatic face
recognition was feasible and to identify the most promising approaches.
Since face recognition algorithms are based on computer vision and machine
learning techniques, measuring performance of face recognition algorithms
requires testing them on standardized data sets. This paper showed that
automatic face recognition was feasible and provided a benchmark for
measuring progress in automatic face recognition. The benchmarks in this
paper became a de facto standard in the research community.
Would you share some of the major advances this
field has seen over the past decade, as well as the problems that
continue to need work?
One major advance in face recognition has been the development of graphics
techniques for manipulating images of faces. These techniques include
morphable models that allow a non-frontal image of a face to be transformed
to a frontal image. A second area of progress is automatic processing of
faces in video. With the availability of inexpensive video cameras that are
designed for uploading video to the web, I see recognition and processing
of unconstrained video of faces as an important and challenging problem.
One of your more recent papers, the 2007 IEEE T
Pattern Analysis paper, "Face recognition algorithms surpass
humans matching faces over changes in illumination," shows that
computer-based face recognition programs actually outperformed humans.
Would you talk a bit about this paper and how the computers won
out?
"Since face recognition algorithms
are based on computer vision and machine
learning techniques, measuring performance of
face recognition algorithms requires testing
them on standardized data
sets."
Many proposed face recognition applications will either replace or assist
security guards. So, a natural baseline for algorithms is human performance
on unfamiliar faces. The results in this paper showed that computers are
capable of out performing people on recognition of frontal face images.
Since 1993, there has been a significant effort to develop recognition
algorithms from frontal face images. Computers won because of a decade's
worth of research in face recognition from frontal images.
What are your hopes for progress in facial
recognition research over the next decade?
I would like to see greater collaboration among psychologist,
neuroscientists, and automatic face recognition researchers. I believe that
by incorporating algorithm models into psychological and neuroscience
studies it is possible to gain greater understanding into how humans
process faces. Also, an understanding of human performance on face
recognition can provide performance goals for algorithms.
What would you like the "take-away lesson" about
your research to be?
I have two "take-away lessons." First, to paraphrase my adviser, "Only work
on problems that are interesting." The second is to collaborate with
researchers in aligned fields. The work on comparing humans and computes
was only possible as a collaborative effort with Prof. Alice O’Toole,
of the University of Texas at Dallas.
P. Jonathon Phillips, Ph.D.
National Institute of Standards and Technology (NIST)
Gaithersburg, MD, USA