Progress in the field of computer vision throughout the past five years has been tremendous. Measured by its impact on products and the economy, computer vision has matured from a nascent set of interesting algorithms and open problems to a real powerhouse. Despite the fact that some of my colleagues like to point out my departure from the field so carefully coincides with this acceleration of success, I’m sure the real reason for this change is that computer vision is now being driven by consumer applications. Our need to recognize, catalog and search visual data has never been greater (or more entertaining). When the consumer market begins to drive innovation, especially in a field with such a solid academic foundation, great (and sometimes scary) things can happen quickly.
Take for example recent work that can search for images of you in very large databases and is smart enough to recognize you based on who you are with in the image. Using classical face recognition methods, the number of operations that are required to find an instance of you is directly related to the number of images searched. Even if computer vision researchers overcome challenging problems like how to recognize you if your face is in shadow, partly occluded, or very far away – they still need to sift through massive amounts of data. You still had some anonymity as a face in the crowd – until now. Researchers at the University of Toronto have developed a method that makes use of your tagging relationships to narrow the search for you. The system builds a network of relationships between you and known individuals by analyzing these tags. Later, when trying to recognize if you are in a photo, these relationships are taken into account to influence recognition. The technique is known as “relational social image search “ (see the lab’s homepage for more details).
The concept is simple: the tags “teach” the system about who you are likely to appear near in photographs. When you appear deep in shadow next to a set of folks you are likely to hang out with, the system can more accurately guess the individual in question is you. The work behind the concept (like most of computer vision) is rich with interesting mathematics and makes for interesting reading.
A patent is pending on the work, and it will be awarded later this month. The researchers, Parham Aarabi, a professor in the Edward S. Rogers Sr. department of electrical & computer engineering, and his former graduate student, Ron Appel, will present the work at this month’s IEEE International Symposium on Multimedia. If you’re looking to expand your horizons from traditional AV or visual computing – it’s a great conference. Maybe I’ll see you there.