Current Issue
This Month's Print Issue

Follow Fast Company

We’ll come to you.

2 minute read

Hipsters Who Won't Admit To Being A Hipster: This Software Could Call You Out

Humans are experts at making snap judgments about other people. A group of researchers has created an algorithm that can do all that judging for you.

Hipsters Who Won't Admit To Being A Hipster: This Software Could Call You Out

How do you spot a hipster? Or a biker, or a surfer, or a goth? Many times, people just "know." Of course, any "knowing" is likely a fashion-based assessment based on a subjective, socially constructed stereotype. But still. You just know. It’s called a snap judgement, and anyone who lives in a big city or went to high school, for that matter, is probably familiar.

Now that millions of photos of people are uploaded to Facebook every day, a team of University of California, San Diego researchers figured that computers should be able to do a similar task: classifying groups of people into "urban tribes." They developed an algorithm that does just that. At 48% accuracy, their program is better than random but not as good as a human yet—apparently it’s pretty hard to train a computer to make a knee-jerk judgment based on subtle cues in clothing and hairstyle.

There could be many valid uses for such software, as the authors note, such as generating more relevant advertisements and content recommendations, improving image search results, and analyzing footage from surveillance cameras. That’s not all necessarily good: Where once the Internet was an open forum for diverse and unexpected interactions, increasingly everything people see is personalized to what or who they already are interested in, even on Google. Algorithms like this one could ensure that, among the eight subcultures included, never shall a leather-donned Hell’s Angel and a skinny-jeaned, fixie rider meet.

The program analyzes groups of individuals, rather than single pictures, which is one reason why it’s performed better than previous efforts. It divides each person into six sections, and analyzes the picture to determine attributes of hair style, head gear, hair color, jewelry, and tattoos. To train the program, the researchers looked on Wikipedia to select the most popular eight "subculture" categories (biker, country, Goth, heavy metal, hip hop, hipster, raver, and surfer) and included photos from self-identified social venues, such as a biker bar or a hip-hop show. They tested the program later on random, unlabeled images, but have yet to test it against an actual human’s own judgment, according to a press release.

"Preliminary experimental results demonstrate our ability to categorized group photos in a socially meaningful manner," they write. The work was presented at the British Machine Vision Conference last year, and they’ve also released an "Urban Tribes" dataset of images, so other researchers can use them. They are working to improve their own results.