Whether it's fashion, a favorite football team, or a certain kind of music, humans seem to enjoy being considered part of a larger group, and often self-identify as such. With this in mind, students from the University of California, San Diego (UCSD) Jacobs school of Engineering are currently developing a computer algorithm that can deduce from an image whether you're a goth, surfer, hipster, or biker.
Since websites such as Facebook have become ubiquitous, a huge number of photos have popped up online, offering new opportunities for the computer vision community, advertisers, and, regrettably, weirdos in general. The students at UCSD Jacobs turned their attention to photos of people at social events in a bid to determine what can be learned, automatically, about the subculture which the people in question belong to.
Using a group photo as reference, the algorithm harnesses a freely available open-source program to separate each person into six separate parts, splitting the face, the main part of the head, the top of the head, neck, torso, and arms into more manageable sections.
The algorithm then considers issues like the normalized average face skin color detected, the type of clothing worn by the subjects, and the top three dominant values in Red, Green, Blue, Hue and Luminance color bands. This latter factor helps pinpoint colors in clothes and accessories, thought to be specific in some urban tribes. Finally, the algorithm analyzes the specific poses of the subjects.
The UCSD Jacobs team cite the potential of such an algorithm in improving recommendation services and gearing advertising to the correct group. The researchers further posit that an individual’s social identity can be defined by that particular individual’s association with various social groups. This seems like an obvious but sound insight, yet offers a chilling glimpse into the potential for government agencies to seek out political dissidents and the like, should the tech be used nefariously.
Still, you can breathe easy as it's early days yet, and the software can only offer an accuracy of around 48 percent – still way above random chance (nine percent) – but a long way off the kind of level that a human operator could offer.
“This is a first step,” explained Serge Belongie, computer science professor at USCD Jacobs, and co-author of the study. “We are scratching the surface to figure out what the signals are.”
Source: USCD Jacobs