You stare at the TV and you wonder: is that character actor in the vintage movie really who I think he is? Dr. Dirk Colbry uses high performance computing to determine if a given face matches the person you think it might be.
Dr. Colbry recently finished his PhD in computer science at
“We work with extremely large datasets. It turns out the work benefits greatly from parallel computing.” In other words, the computational problem is well-suited to being broken up into small chunks that can be divided among a number of processors, or CPUs. “My algorithms are perfectly suited to the HPCC” he says. “If I used my approach on the fastest PC in the world, it would take 1 ½ years to complete. The HPCC can solve most of my simulations in a day or two.”
When Dr. Colbry says “large” he means it. In one test he performed 5000 facial scans. He used a 3D camera to capture images. Each of those scans produced files 3 megabytes in size. To exhaustively perform all pairwise comparisons, 25 million operations are required. Comparing 25 million images, each 3 megs in size, requires some serious computational horsepower.
Dr. Colbry explains that most conventional work in facial recognition analyzes only two dimensional data. This deprives the computer of an entire dimension of “depth” information. While the 3D approach requires much more computational effort, it yields much more precise results and is tolerant to variations such as facial pose and ambient illuminations.
Dr. Colbry notes an important distinction in the field of facial recognition: While “verification” seeks to determine “Is this an image of the person we think it is?” “recognition” seeks to match a given image against a large database of candidate images. While science fiction movies may portray things differently, it’s much harder to compare an image against thousands or millions of candidates than it is to verify that a given image is that of a specific subject.
In either case, the task can be daunting. “Often we work with images that lack the quality we need. Lighting may be poor. A subject might have grown a beard. And if we only have 2D facial images we have so much less information available to work with.”
It is somewhat eerie to observe a wireframe image of a human face and to visualize how facial recognition software seeks to match that image with a real human. The software literally captures the outline of the face, so as to compare that abstract construct with the thousands of candidate matches.
Facial recognition technology obviously plays an important role in forensics and in homeland security. The National Institute of Standards and Technology (NIST) sponsors a Facial Recognition Grand Challenge (FRGC). Dr. Colbry notes with a smile that some of his research at
Do human faces resemble tire treads? Perhaps not, but Dr. Colbry and colleagues have applied similar pattern recognition techniques to solve the question of whether a given set of tire track images matches an entry in the database of tire track images.
Dr. Colbry collaborated with MSU professor George Stockman and university distinguished professor Dr. Anil Jain on his research into pattern recognition. Dr. Jain leads MSU’s biometrics research efforts. The biometrics lab studies various ways that pattern recognition can analyze faces, tire treads, fingerprints, or other patterns in nature or in human endeavor. See http://biometrics.cse.msu.edu.
Dr. Colbry joined
