Looking forward to a substitute for human vision
One of the greatest mysteries that hide in plain sight is human vision. While psycho-physicists and neuroscientists investigate the real thing, computer science has looked at it another way, by trying to make machines that can 'see'. Professor Rama Chellappa from the University of Maryland is at the forefront of this research. He has gone beyond simple pattern recognition, into a new area that tries to find ways of extracting useful information from an image.
Along the way, applications are emerging where 'seeing machines' can replace humans in hostile environments such as space and the battlefield and, much closer to home, can help in security and intelligence work by patiently looking at crowds of people to spot anyone from a shoplifter to a terrorist.
Professor Joseph JaJa, director of the University of Maryland's Institute for Advanced Computer Studies, said: 'He is a real gentleman; that captures his character. He is very considerate, a team player, works extremely well with all sorts of people - easy-going ones and difficult ones. Of the 75 people in my faculty, he is one of the top one or two in terms of being a statesman and a gentleman in that group. If his students get sick, he goes out of his way to help them; as a person, he is absolutely wonderful.
'He is in the top three or four people in computational image processing. He is very well known in two separate communities who usually do not talk to each other: the image-processing community and the computer-vision community. He has done seminal work in both areas, and builds bridges between the two communities. The work that received the most attention, and has been extremely well cited, is concerned with motion recovery. He had a very novel approach to solving this problem and he is a pioneer. He has also pioneered the Markov Random Field Model for texture analysis and synthesis. He developed the model, and now there is a whole school that is following this approach. It is a part of image processing, but this approach started with him.'
'Another area that has got a lot of recent attention is human face and gait identification and, a couple of years ago, independent tests rated his algorithms in the top three.'
'There are two people I rank as technically the best in the institute, but Rama is much broader. We made him a very good offer to attract him here from USC and this is probably one of the best things that we have done in the past 10 years.'
Rama Chellappa was born into a middle-class family in Madras, India. His father had qualified as a teacher and loved his work, but discovered that he could not make enough money doing it to raise his family, so he took a better-paid job with a bank. Chellappa was always interested in science, but decided to study engineering at Madras University because he thought it would be more useful and lead to a better job. He did well enough to gain a post-graduate place at the prestigious Indian Institute of Science in Bangalore - which subsequently became the focus of India's own 'Silicon Valley'.
After gaining a master's degree, Chellappa had to decide whether to continue his education and enter research in India, or to travel abroad.
He said: 'I had done some courses in image and pattern recognition, and had decided that this was a field I wanted to work in. At that time, research in this area was more advanced in the US. Also, if I really wanted to go back to India and get a faculty position, it would be better to have a PhD from a US university. I was not totally sure at that time where I would be going, but getting a US PhD gave me more options.
'I had offers from several universities, but at that time Purdue University was the leader in pattern recognition.'
After obtaining his master's degree at Purdue, Chellappa decided to move to the University of Maryland to work with Professor Azriel Rosenfeld, a seminal figure in the image-processing world, to do a PhD in computer science. But after six months, he changed his mind and returned to Purdue to do a PhD in Electrical Engineering under the guidance of Professor Kashyap, who is well known for discovering the Ho-Kashyap algorithm. He kept in touch with Rosenfeld and he had a part-time appointment in Rosenfeld's lab, where he worked during summer and winter breaks.
Chellappa said: 'At Purdue I was exposed to work on statistical pattern recognition from Professor Ken Fukunaga, who wrote the most influential book on the subject. He was also a brilliant teacher. I was also fortunate to have learned syntactic pattern recognition from Professor K.S. Fu, who literally invented that field. My advisor, Professor Kashyap, well known for his work on statistical and stochastic methods, taught me the principles of statistical inference and decision theory. I was also fortunate enough to be chosen to proofread a book by Professor Jack Sklansky, from the University of California at Irvine, and I managed to make a very small contribution to the book and so Jack became one of my mentors. Also at Purdue at the time, there was a Professor Thomas Huang, who was teaching a course on image processing. I was exposed to what were really some top-class researchers, and I think that has helped me.'
Having gained his PhD, Chellappa was looking for a job where he could continue working in image processing, and was lucky to find an opening at the University of Southern California which, amongst other things, gave him a break from the Mid-West weather.
Chellappa said: 'USC had an image-processing institute, which had been started in the early 70s by two pioneers, Bill Pratt and Harry Andrews. Both had then left USC for industry positions in the 1970s. Although I was a very young professor, I thought I could grow into a good position there, so it was the best place that I could have gone to after graduation.'
During his time at USC, Chellappa changed the direction of his research from the problem of just modelling image data to something that he felt would be more use in the real world. That is, trying to get machines to recognise three-dimensional objects from the world - to give machines something close to human vision.
He said: 'In computer vision, we try to extract three-dimensional descriptions from one or more two-dimensional images, so this becomes a very interesting inference problem. It seems more challenging, because it combines statistical inference with geometric inference. When you are trying to infer things about a 3-D scene, you have to worry about the geometry. Since you are doing it from data, it becomes an inference problem. It throws up a lot of interesting theoretical problems. A lot of the time, the problems are ill-posed, so you have to worry about coming up with robust algorithms. Humans are good at solving these problems, and so we are looking at ways that machines can do the same thing.'
After 10 years in Los Angeles, Chellappa was finding life a bit 'hectic' and, with a young child, felt it was not a great place to bring up his family. He started looking around for alternatives, and mentioned to Azriel Rosenfeld that he was thinking of moving. Rosenfeld immediately asked him to come to Maryland. He said: 'USC was fine - it is one of the best universities in the country - but I was looking at things like schools and the kind of home I could afford to buy. I really miss the beach though.'
Chellappa has achieved a high profile at Maryland. His research on machine vision has attracted a lot of attention from the commercial world, and he has become director of the Center for Automation Research. His work has also come to the attention of those who study security. He has been working on ways that machines can recognise people from video images, both by their faces and by their gait. This appeals to those who believe that, one day, it may be possible to put cameras at airports, and even shop entrances, to be able to spot known criminals as they enter the country - or shoplifters before they enter stores.
Chellappa believes that we are still a long way away from a reliable system to do this. He thinks that the best systems today, working in idea conditions, are still no more than 50 per cent accurate. There are too many variables from lighting, to angle, to whether someone is moving away or towards the camera, or how much they have aged since the reference picture was taken. This is still something that humans can do very easily and much more accurately. But his work has featured in a National Geographic article on biometrics and homeland security.
Chellappa is very much driven by the applications for his research. He believes that machines can be made that can go into environments hostile to humans and perform tasks autonomously. But he accepts that we are a long way from even understanding human vision, and that it will be generations before we get there. He said: 'There are a lot of people working on human vision, and our job is to understand as much of that as possible and try to make machines that can work reasonably well. People who work with human vision attempt to derive computational models of perception, and our goal is to see if machines can be built to do the same. For example, one of the things we have found recently is that, when humans look at a 3-D scene, there is a definite bias as to how they judge the depth of points in that scene. We have been able to quantify what the bias will be, depending on how the object or the observer is moving. The human-subject researchers have observed it, but we have come up with an explicit expression to describe it.
'My modest aim is to try to understand what is going on in human vision research, and then try to replicate some of its performance.'
Chellappa has got so used to living in the US that he no longer thinks of returning to India. He keeps in touch with his cultural roots by performing South Indian Carnatic singing. Every year, he joins other ex-pats for a festival in Cleveland. One of his brothers has also moved to the US and works in finance in Wall Street. His other brother, who is also in finance, is living in India.
University of Madras, B.E.
Indian Institute of Science, Bangalore, M.E.
Purdue University, M.S.E.E. then PhD
Assistant Professor, University of Southern California, Los Angeles
Associate Professor USC
University of Maryland, Professor of Electrical Engineering and affiliate Professor of Computer Science, Director of the Center for Automation Research and Permanent Member of Institute for Advanced Computer Studies