Admit it, you've done this: You've stereotyped the mustachioed guy in thick-rimmed glasses as a "hipster," the girl in excessive black eyeliner as a "goth," the beefy guy in leather as a "biker." These terms are all admittedly obnoxious. But the fact that we often wear our social identity in our accessories speaks to the reality that cities are made up of myriad subcultures. Sociologists call them "urban tribes."
Obviously, it doesn't take a lot of nuance to identify these groups with your own eyes. Identifying them by computer, though, is a much trickier task. Computer Scientists at the University of California at San Diego, Columbia University and the University of Zaragoza in Spain have actually been developing algorithms to do this – to automatically mine group photos for markers of collective identities.
Why, you ask, would we want a computer to do that? As the researchers argue in a paper [PDF] explaining their process, algorithms capable of ID'ing group identities from photography might inform smarter image search engines, targeted advertising, or more accurate recommendations on social networks.
Maybe you don't actually want Facebook inferring from your party photos that you might like this ad for organic quinoa, or this local listing for an Ender's Game book club. But it's weirdly fascinating to know that this might be possible.
The researchers built the project around the eight most common subculture categories listed on Wikipedia: biker, hipster, country, goth, heavy-metal, hip-hop, raver, and surfer (which brings up the point that, yeah, a lot of us don't fit into a clear subculture at all). They also searched for photos on various search engines with social tags like "club," "formal" and "pub."
Each of the photos used in the project was automatically disassembled into its individual human parts, including the face, head, neck, torso and arms of each body in the image. The researchers then systematically trained the algorithms to detect evidence from those details – a tattoo here, a certain kind of hat there – to piece together a social identity for the whole picture.
At this point, their technique is accurate 48 percent of the time, which probably isn't high enough to interest advertisers quite yet. But, hey, it's a fair amount better than random chance (that would be closer to 9 percent).
Top image courtesy of the Jacobs School of Engineering/UC San Diego.