Get to Know Chellappa

I am very fortunate to have been taught by the best in the fields of Pattern Recognition and Computer Vision. Their formal and informal teaching and guidance have had great impact on how my academic career has developed. In this brief article, I describe my interactions with my teachers while I was a graduate student at Purdue University and the University of Maryland as well as my early research and how it incorporated their influence and guidance. In doing so, I hope to shed some light on the brilliance, innovative spirit, work ethic and generosity of my teachers who also happen to be the pioneers in these fields.

I was introduced to statistical pattern recognition in Spring 1977, when I took a course on this topic at the Indian Institute of Science (IISc). Not knowing much about the topic, I took this course because the instructor Prof. M.A.L. Thathachar was known as an outstanding scholar and teacher. My interest in pattern recognition was sparked by the well delivered lectures of Prof. Thathachar; and as a result, I got interested in doing my doctoral studies in pattern recognition. If one were interested in pattern recognition, the Electrical Engineering Department (as it was known then) at Purdue University was the place to be, in those days. The faculty roster included Profs. R. L. Kashyap, King-Sun Fu, Ken Fukunaga, Tom Huang, and E.A. Patrick.

I was fortunate to get a research assistantship from Prof. R. L. Kashyap to pursue my doctoral studies. I am singularly fortunate to have taken classes in pattern recognition from Profs. Kashyap, Fu, and Fukunaga. In Fall 1977, I took a course titled Introduction to Decision and Control Under Uncertainty from Kashyap, which covered time series models, parameter estimation, Bayesian information criterion, and stochastic filtering. Kashyap was at his best while teaching Bayesian inference. This course laid the foundations for my doctoral work on 2-D Markov random field (MRF) and non-causal autoregressive models. The love for MRFs has not left me and my students, and we seem to be coming back to them on and off.

In Spring 1978, I took the course, titled Artificial Intelligence, from Prof. Fu. The course material was equally divided among statistical pattern recognition, syntactic pattern recognition, and introductory material on artificial intelligence. Artificial Intelligence (AI) was a new field then and teachers liked to include AI in pattern recognition courses. Prof. Fu was a dynamic teacher. I also audited a course on image processing taught by Prof. Tom Huang in Spring 1978. Prof. Huang came to the class carrying a mug of coffee and without any notes. His lectures were enjoyable and quite technical at the same time.

In Fall 1978, I took a statistical pattern recognition course from Prof. Fukunaga, who used to bring his book and put it on the table, put a piece of chalk in a silver holder and write almost the entire book on the board! Prof. Fukunaga’s course had the official title, Introduction to Artificial Intelligence, but after the first introductory lecture, it was all about statistical pattern recognition. What I remember from his course is that he gave no homework, but gave several tests! Prof. Fukunaga’s gift was in teaching even complicated concepts such as deriving the probability of error for general non-linear classifiers in a simple manner. It was apparent that he loved estimating errors bounds for different classifiers and somehow made these derivations look easier!

Prof. Fu also taught a course on Introduction to Formal Languages in Fall 1978, which introduced finite-state, context-free and context-sensitive grammars, their parsers, stochastic grammars and how to develop syntactic pattern recognition systems. He was fond of discussing the challenges in inferring the production rules for a given pattern recognition problem. His passion for structural and syntactic methods came through at every lecture. Despite a heavy travel schedule, Prof. Fu rarely missed his lectures.

During the 3.5 years I spent at Purdue, I also took courses on mathematical statistics, multivariate analysis, a seminar on multivariate analysis (all taught by the veteran statistician Prof. K.C.S. Pillai). In addition, I took a course on non-parametric methods and audited a course on discrete decision theory and an advanced version of it taught by Prof. Jim Berger. The Purdue statistics department was quite strong on decision theory in those days with the likes of Prof. Shanti Gupta, Jim Berger and Herman Rubin on the faculty. I recall a symposium on decision theory held at Purdue in the summer of 1981. Prof. C.R. Rao was an honored guest and gave a fantastic talk, and I was the honored Chauffer who gave him a ride from a party to the Purdue Memorial Union where he was staying, and in return he signed my copy of his book on Linear Statistical Inference. I was awed by the reverence everyone at the meeting showed to Prof. C.R. Rao.

After receiving a Master’s degree in EE in December 1978, I joined Prof. Azriel Rosenfeld’s group at UMD to pursue computer vision research and get a Ph.D. in Computer Science. After a semester at UMD, I decided to return to Purdue in Summer 1979 to get my Ph.D. in EE under the supervision of Prof. Kashyap. Azriel, being the gentleman he was, continued to support me as a Faculty Research Assistant at UMD, while I was a Graduate Research Assistant and Instructor at Purdue. Azriel served on my Ph.D. proposal and dissertation committees and visited Purdue to attend these exams. I visited his lab during winter and summer breaks and wrote several reports published by the Computer Vision Laboratory at UMD. The day of my defense was exciting as Profs. Kashyap, Fu and Rosenfeld invited me to join them for lunch before the exam. Azriel gave a talk at Purdue after my exam using just one transparency foil!

In Fall 1977 and Spring 1978, I worked on the Bayesian model order selection problem for autoregressive and moving average models (ARMA), extending Kashyap’s work on Bayesian comparison of time series models which appeared in IEEE Transactions on Automatic Control in 1977, several months before the work of G. Schwartz which appeared in Annals of Statistics in 1978. One of the results I learned for handling the messy likelihood function of an ARMA model was the Gaussian approximation of the ARMA likelihood function which is referred to as Laplace’s approximation result. Kashyap made me read the book on asymptotic expansion by Erdelyi which contains many approximations that can be useful. During my doctoral work on stochastic models for image processing and analysis, Kashyap’s guidance was invaluable in deriving parameter estimation and neighborhood selection rules for 2-D MRF models and 2-D non-causal spatial models. We were all excited by the papers of famous statisticians such as P. Whittle (Biometrika, 1954) and J. Besag (JRS-B, 1974) and explored the applications of their work for texture analysis, image restoration, and classification. Although MRF models are not being used much for synthesis and analysis problems, the inference methods inspired by MRF models are being used even today in computer vision. As an example, one can point to the recent work known as Make 3D and a face alignment algorithm developed by one of my students.

At the Image Modeling workshop organized by Azriel in June 1979, Prof. Fu introduced me to a representative from Springer, who was looking for a graduate student to go over the manuscript on Pattern Classifiers and Trainable Machines by Profs. Jack Sklansky and Gustav Wassel. I had the pleasure of reading this book before it was printed. I offered to write a brief section on the role of sufficient statistics and Bayesian learning. Jack generously agreed to include this section in his book. This fine gesture by one of the leading researchers taught me how to encourage and mentor my students and younger colleagues.

While I appreciate the high quality instruction I received from the pioneers mentioned above, what I truly appreciate is the impact Profs. Kashyap and Rosenfeld had on me as I worked very closely with them. Kashyap took it upon himself to improve my analytical skills and the quality of my technical writing. He introduced me to many classical books on decision theory and made me write every paper I wrote with him numerous times. He was never tired of editing my papers and suggesting improvements. The first position paper on optimal decision rules I wrote for him in Fall 1977 was edited more than ten times. At one of meetings in the first few weeks of my joining Purdue, Kashyap suggested I buy Scotch tape, scissors and a mechanical pencil. As each revision was done, there would be more and more cutting and pasting from earlier versions, so that the entire paper need not be rewritten for the next reading. Thus, as the iterations proceeded, one had more to cut and paste and less to add. This was Kashyap’s way of instilling a passion for technical writing which is a must for a successful academic career.

I spent nearly fifteen years with Azriel, first as his student in Spring 1979, and then as a faculty research assistant during August 1979 - August 1981 while I was a graduate research assistant at Purdue. I also spent an additional 12.5 years as his colleague at UMD (during 1991-2004). Azriel was one of the hardest working and most brilliant professors I have known. He rarely took vacations and spent close to 16 hours/day on writing, reading and thinking, and lecturing computer vision and other subjects that interested him. He was such a prolific writer; he wrote a technical report while we were on an excursion trip from Bangalore to Mysore (~110 miles) in India in January 1988. He published this work a bit later. He also patiently edited all the papers written by his students, colleagues, and visitors. Every night, he used to carry home many papers, which he would edit till late at night and bring the edited versions the following day. If for some reason, he could not complete editing all of them, he was always very apologetic. Despite managing several large grants and managing a Center, he regularly met his students and advised them. He was known for taking red eye flights from the west coast often and coming to work for a full day of work. Traveling with him to conferences and other meetings was a blast as he had a good sense of humor. Above all he took immense pleasure in mentoring young researchers and making them better in what they do. The computer vision field benefitted enormously by his vision, work ethic, and integrity.

Although my interactions with Prof. Fu were mostly through his classes and his participation as a member of my MS and Ph.D. committees, he had an enormous influence on me and other students as well as numerous colleagues. When he was in town, you could see him in his office poring over his papers late at night. As a result, most of us who were students in the Advanced Automation Research Laboratory directed by him were also motivated to work nights and weekends.

My interactions with Prof. Tom Huang deepened when I got interested in structure from motion problems in the mid eighties. Since then, Tom has been a great mentor to me and many other researchers. His continued passion for research and scholarship is an inspiration to all of us.

I feel the training I received from the pioneers mentioned above during my years as a student has helped me a great deal to accomplish whatever little I have been able to do. They taught me how to teach, do research, and advise and mentor students. I am truly grateful to them.

I am also fortunate to enjoy collegial interactions with many leading researchers in image processing, computer vision and pattern recognition. In addition to being taught by the best, I have been very lucky in getting several tens of outstanding students who, over the years have tirelessly taught me all the new things they were working on. I just cleverly positioned myself between outstanding teachers and smart students, and it has been a great ride. My only regret is that I have not had an opportunity to study under Prof. Ulf Grenander, who pioneered pattern theory, or Prof. David Mumford, who brought rigorous mathematics to computer vision, even for a summer. This would have completed my career. I do read their books on and off and hopefully one day I will understand some of their research findings.