Team takes top rankings in computer vision ‘Olympics’
For two months last fall, algorithms developed by computer scientists throughout the world quietly worked to detect and classify objects as part of the International Large Scale Visual Recognition Challenge 2013 (ILSVRC2013).
The Challenge required teams’ algorithms to systematically search through 40,000 images in an attempt to identify 200 objects. When the virtual dust settled, the NEC-MU team had taken second place worldwide, claiming top honors among U.S. teams.
Associate Electrical and Computer Engineering Professor Tony Han and doctoral candidate Miao Sun made up the MU half of the team. Xiaoyu Wang, mentored by Han to earn his doctorate at MU in 2012 and now working at NEC — a multinational provider of IT services and products in Japan — rounded out the team.
“These algorithms are like giving a computer real eyes,” said Han of the work, explaining that before the Challenge, participants received several thousand training images along with ground truth solutions to test their techniques.
NEC-MU team’s core technique, “Regionlet for Generic Object Detection,” has drawn lots of attentions in the computer vision community.
“The ultimate goal of object detection is to let a machine know every object it sees, from a tiny pin to a person,” said Wang who has been working in this field since 2005.
“By parsing the fundamental objects in an image, researchers can pursue further goals such as scene understanding, action recognition, social activity analysis and more. The wide application of object detection as well as its importance in computer vision interests me to work on this hard core problem,” Wang said of his interest in object detection.
Originally, researchers in Han’s lab and those at NEC intended to enter the ILSVRC2013 separately, but ultimately teamed up to take the Challenge.
“We found our algorithms were complementary, so we decided to work together. We combined them to submit the results,” said Sun who did an internship at NEC in preparation for the Challenge. “Xiaoyu gave me some instructions to improve Mizzou’s algorithm.
“Computer vision is a very hot topic, not only in universities but also in industry,” Sun added. “Automatic driving and Google glasses need computer vision.”
Han, who is very excited about the team’s performance, refers to the Challenge as the “Olympic game for computer vision and machine learning researchers.” He is quick to point out that the competition included teams from much larger schools and organizations such as the University of California-Berkeley, Microsoft research, IBM research, New York University and Oxford University, among others.
A team comprised of researchers from the University of Amsterdam and Euvision Technologies was the only group to score higher than the NEC-MU team.
Besides increasing the research visibility of Han’s lab, the prestigious win bodes well for Han’s graduate students’ employment futures.
“Web giants such as Microsoft, Google, Facebook or even startup companies like Pinterest are putting a lot of resources on this topic,” said Wang.
Sponsors for ILSVRC2013 were Stanford University, the University of North Carolina-Chapel Hill and Google.