For over a decade, computer vision has been on the fringes of computer science. Neither computer vision, nor supporting subjects such as graphics and geometry, are reliably present in the ``core'' CS curriculum. Our work has required expensive specialty equipment, not used by other CS researchers, and demanded efficiency to a degree that has precluded using many modern programming tools. Specialization within artificial intelligence has created barriers between computer vision and other areas of AI (robotics, ``core'' AI, natural language).
This situation is about to change, I plan to be on the forefront of that change, and my background makes me ideally suited to be there.
The driving force behind the upcoming change is the increasing speed and decreasing cost of hardware. A decade ago, even top research labs had equipment only barely suitable for computer vision research: small disks, slow processors, 1-bit graphics screens, and high-noise, black-and-white cameras. This equipment was expensive and bulky. Locally-built boards were not uncommon. Researchers had access to only tiny collections of images.
Imaging is now becoming mainstream. Even laptops now have color display screens, large disks, and sufficient processing power to run computer vision algorithms. Large collections of images are sold to the home PC market on CD-ROM. Many homes have scanners. Color cameras for the PC are now the size of 35mm cameras and rapidly dropping towards the price of other common peripherals. Stores sell a variety of shrink-wrap packages for graphics and image processing, as well as cookbooks for coding such algorithms. Now everyone can do imaging, cheaply and in bulk.
Adapting to these changes requires re-thinking how we do research in computer vision, and how computer vision relates to the rest of computer science. Computer vision researchers must learn to work with the new, PC-based equipment. We must rethink our experimental methods, to handle much larger and diverse sets of test images, and to exploit the higher quality of these images. We must impose on ourselves a higher standard of software engineering and scientific documentation, to create a working development path from research prototypes to reliable products.
We must also ask how researchers in other areas might use images. Computer vision researchers are already starting to collaborate with experts in database design to handle large libraries of images. If images become ubiquitous, programming languages should contain proper support for them. Geometrical and graphical algorithms should become part of the programming and algorithms curriculum. Fast transmission of images should be incorporated into architecture and networking, not side-lined into a ``multi-media'' ghetto. Finally, it is time to reopen the long-standing question of how high-level vision can be connected to artificial intelligence and natural language semantics.
I am well suited to help lead the field through this transition, because my background is very broad. My research career actually began in linguistics and computational linguistics, with side-interests in other areas of AI and in LISP language development. I have a basic familiarity with most areas of mathematics, with a deep knowledge of low-dimensional topology. (I often attended the topology seminar at U. Iowa.) My work in computer vision has required me to become familiar with the basics of statistics and optics. Together with my graduate students, I have had to assemble camera mounts, calibrate lenses, debug hardware failures with our robot arm, and install and maintain the operating system (linux) on our laptops.
My current research is a mixture of pure computer vision and projects which relate computer vision to other areas. These include:
Related practical projects include:
I have previously worked in the following areas:
This page is maintained by Margaret Fleck.