Associate Professor of Informatics and Computing
A Vision for Computer Vision
Sometimes little things can open a world of imagination. For Associate Professor of Informatics and Computing David Crandall, it was a Sinclair ZX81 – a simple computer with no moving parts and only 1Kb of memory – that inspired him to study computer science, eventually putting him on the cutting edge of computer vision research.
“My parents, from very early in the digital revolution, had computers at home,” says Crandall as a slight smile crosses his lips. “The ZX81 is really funny to look at now, 30 years later. It is so antiquated. But when I was very little, maybe five or six, it really inspired me to learn how to program a computer.”
Crandall’s love of computers and programming didn’t fade, and by the time he entered high school, the computer revolution was in full swing. He immersed himself in computers, reveling in taking them apart and putting them together, getting to the guts of both the software and the hardware, to learn how they worked from the inside out.
“Everything was very open then,” Crandall says. “You could open a computer, pull things out, wire in new circuits, and see what happened. The way I approach problems these days has that same tinkering kind of aspect.”He went on to Penn State University to study computer science and engineering, and became involved in the computer vision lab there as an undergraduate research assistant. After finishing a joint B.S. and M.S. in 2001, he worked as a research scientist at Eastman Kodak for a couple of years. He went back to school and earned a Ph.D. in Computer Science from Cornell University in 2008.
He arrived in Bloomington in 2010, and now directs the IU Computer Vision Lab, which employs advanced machine learning and statistical techniques to extract and understand the semantic information stored in images. In other words, Crandall is trying to teach computers how to see.
“Computer vision is really a hard problem,” Crandall says. “The very first people who worked on it in the 1960s literally thought they would be able to solve it in a summer. People can see -- why can’t computers? It turns out it’s very, very difficult. After so many years of work, my dog can still identify objects better than the fastest supercomputers in the world.”
But progress has been accelerating, thanks to a number of breakthroughs in the last few years. Perhaps the most important one is also the simplest: digital photography has become so popular that vision researchers can access many more images than ever before. Crandall and his colleagues have been mining images from photo-sharing websites such as Facebook and Flickr to expand the library of information computers can draw from when they’re learning to recognize images.
“Only 10 years ago, the way we collected images to develop our algorithms was to walk around campus and take a few pictures,” Crandall says. “Now we can download millions of public images from the web instead. That means we can feed our machine learning algorithms a much larger and more representative sample of what the world actually looks like.”
Crandall and his students work on a variety of computer vision problems, from teaching computers to recognize objects like people, cars, and birds, to automatically building 3D models of cities using images downloaded from the web. One of their goals is for computers to someday recognize images well enough to make it easy to search for a photo just by describing it to the computer.
“With tens of thousands of photos on my computer, it can be really hard to find the one I’m looking for – I often have to page through each one individually.” Crandall says. “Someday I’ll just be able to tell the computer `find that photo of me standing next to my grandmother on her front porch,’ and it will understand exactly what I mean and do that for me.”
Besides organizing photos, other applications of their work include assistive devices that could help the visually impaired and robots that could better perceive the world. They’re also applying their computer vision technology to help scientists process massive image datasets in other domains, like ecology, psychology, and earth science.
“It’s a really exciting time to be working in vision,” Crandall says. “The technology is finally starting to work and big companies are investing a lot of money in this area. I think we’ll see many more breakthroughs in just the next few years.”
With Crandall and his colleagues working on the problem, computers might learn to see well enough to help you organize your photos and so much more.