Being able to separate the famous from the hoi polloi has certain financial implications for YouTube. Stars attract viewers, so being able to identify videos with A-List talent might suggest where to direct marketing and promotional resources. And because stars tend to feature prominently in videos uploaded to YouTube without authorization, being able to identify popular icons could help spot copyright violations.
Such speculation may offer some hint as to why Googles researchers have been working on celebrity facial-recognition technology.
In a paper titled Audiovisual Celebrity Recognition In Unconstrained Web Videos, presented at a conference in Taiwan last month, Google researchers Mehmet Sargin, Hrishikesh Aradhye, Pedro Moreno, and Ming Zhao describe work they ve been doing to recognize the faces of celebrities in videos.
The problem is that facial recognition on video doesn t work very well in many circumstances. The researchers note that successful examples of facial recognition on video tend to involve constraints, such as a TV anchor appearing as a talking head, without much movement, under controlled illumination.
Thus, the researchers have developed an automated technique to associate faces detected in an image with people s names found in Web pages text or tags. The system doesn t need to be taught and can accommodate the scale at which YouTube operates.
The Internet is in a constant state of flux, and new celebrities are constantly added to the popular culture even as the celebrities of the past fade, the paper states. This ability to learn autonomously to constantly add to the existing gallery of celebrities is therefore a major design principle of our work.
The system can recognize hundreds of thousands of celebrity faces by exploiting the tremendous depth of the Internet, according to the paper.