Colorado State University (CSU) Computer Science Professors Ross Beveridge and Bruce Draper did their Ph.D. work in the UMass Amherst Computer Vision Group, under the joint direction of Professors Ed Riseman and Al Hanson, and both graduated with Ph.D.s in 1993. The pair currently codirect the Computer Vision Research Group at CSU, following in the footsteps of Riseman and Hanson, leading the next generation of vision researchers.
At UMass Amherst, Beveridge and Draper saw firsthand an unusual model for academic research: namely, two good friends with very different personalities jointly coordinating a research group with multiple different projects. "At UMass, graduate students in the Vision group quickly learned that they didn't have one advisor, they had two," says Beveridge. "Whether Ed or Al was leading a particular project, the other was always following the work carefully and would quickly jump in with ideas of his own. Graduate students working in the CSU Computer Vision Group today learn a similar lesson, and the group is stronger for it."
Major accomplishments by Beveridge's and Draper's CSU Computer Vision Group include advanced analysis of algorithms that perform human face recognition and real-time activity recognition in video. In face recognition, the group has taken a leading role in the evaluation of algorithms, working closely since 2003 with Dr. Jonathon Phillips at NIST on a series of public face recognition challenge problems. In joint work with Geof Givens in the CSU Statistics Department, they have carried out the most detailed studies of interacting factors that influence face recognition algorithm performance.
Beveridge's and Draper's team also releases and maintains a suite of open source software packages that support labs around the world that want common face recognition baseline algorithm implementations and software for carrying performance studies according to standardized protocols. Most recently the researchers have taken the lead in a competition associated with the 2014 International Joint Conference on Biometrics where different labs and groups will work on a common video face recognition dataset.
Beveridge and Draper also work closely with CSU Mathematics Professors Michael Kirby and Chris Peterson on problems at the intersection of computer vision and algebraic topology. This work blends mathematics with practical utility in some surprising ways, including fundamental advancements in how to approach the problem of recognizing people's actions in video.
Their Vision Group is also known for its work on biomimetic approaches to computer vision, with an emphasis on unsupervised learning. "At its heart, this approach represents a commitment to developing systems that learn important internal representation without detailed guidance by people," says Draper. To express this in practical terms, the common approach to many computer vision tasks today is to require humans to hand label thousands of images, e.g. "That is a cat. That is a truck." This is tedious, and ultimately visually intelligent systems should not require such extraordinary hand-holding.
Recently, Beveridge's and Draper's group's work on biomimetic approaches to vision has been combined with the work on high-dimensional manifolds in the context of real-time activity recognition in video. This work was supported by the DARPA Minds Eye Program. A major component of their work was incorporated into a larger system that was then demonstrated by iRobot. The system learned quickly and then recognized actions such as running, walking, picking objects, carrying objects, etc. A video highlight of this effort is available at: www.cs.colostate.edu/~draper/.
Beveridge, Draper, and their team are now merging the video understanding and face recognition work. More generally, they remain committed to the same goals set out by Riseman and Hanson at UMass Amherst: machines should be using their eyes to watch, should be making sense of what they see, and then should be stepping in to help people in the myriad tasks that arise in all aspects of life.