A new machine-vision strategy for data-mining a massive database of high school yearbook photos has been pioneered by University of California, Berkeley's Shiry Ginosar and colleagues.
More than 150,000 pictures were downloaded from an index of images dating back at least a century, and 37,000 images remained after weeding out non-full-frontal portraits. The portraits were then clustered by decades, and the images superimposed to produce an "average" face for each period. The process extracted other "average" characteristics for each period such as hairstyle, clothing, and facial expressions.
The researchers found that until photography's popularity surged during the 20th century, expressions remained neutral, and afterward smiling came into vogue. An algorithm designed to interpret the degree of lip curvature in the photos found a clear trend in growing smile intensity as time passed, and women consistently grin more than men.
The researchers also observed the evolution of hairstyles over the decades. "By use of a large historical data collection and a simple smile-detector we arrived at the same conclusion with a minimal amount of annotation and virtually no manual effort," the researchers note.
From Technology Review
View Full Article
Abstracts Copyright © 2015 Information Inc., Bethesda, Maryland, USA