MIT Researchers Develop 'Feature Tuning' Algorithm

Apr 25, 2014 | Contact Author
Your message has been sent.
(click to close)
Contact the Author
Save
This item has been saved to your library.
View My Library
(click to close)
Save to My Library
Title: MIT Researchers Develop 'Feature Tuning' Algorithm
algorithmx tunex facex imagesx memorabilityx
  • Article
  • Keywords/Abstract
  • Related Material

Keywords: algorithm | tune | face | images | memorability

Abstract: Researchers at MIT have developed an algorithm for the “feature tuning” of digital images, specifically faces, for improved memorability and other qualities. While the technology is implicated for effects in social media, it also suggests new considerations for digital imaging in cosmetic product development.

In a recent study from MIT, researchers developed an algorithm to assess digital photos for their memorability, along with an approach to adjust or "tune" them in such a way to highlight a person’s most memorable features. This process was detailed in work presented at the International Conference on Computer Vision in Sydney in December 2013, and earned researcher Aditya Khosla a Facebook Graduate Fellowship for further research.

According to a report on Scienceline.org, initially Aude Oliva, a researcher with MIT’s computer vision and graphics group, published work on what makes images memorable. She and colleagues found that memorability is an intrinsic property of images, and input more than 2,200 photos and their "memorability scores" into a computer to teach it how to predict the memorability of photos. According to the report, the photos that ranked highest for memorability were not necessarily of individuals who were more beautiful or had bigger smile. Instead, the computer was picking up on a vast array of tiny details that contributed to memorability.

Khosla then posed an idea to Oliva: If a photo is memorable due to differences in a few pixels, could changing those pixels change the memorability? He and colleagues created an algorithm to capture that memorability and impart it to other photographs. They focused on faces because small changes could greatly impact the photo and the recognition of the subject in the photo. As small changes were made to the photo, the algorithm gave the image a new score for memorability. By measuring 10,000 randomly changed versions, the program pinpointed the most memorable face that still looks like the subject.

Beyond memorability, the algorithm could be used to tune other features like age, attractiveness and friendliness. As Khosla stated, in the report, “If we could pick or modify images to make them extremely memorable, we’d just have to show them once. ...You don’t really have a choice but to store the information because that’s how we are all programmed." One could further envision changing photos of products or packaging in smart ways to create an instant connection with consumers.