Improve your selfie: Researcher can predict a photo's popularity
(Phys.org) —Want to rack up more views for your selfie? We can't all take our photos with the president, but a new tool developed by an MIT PhD can help us predict how popular our photos will be.
A recent study led by Aditya Khosla, a PhD student at MIT's Computer Science and Artificial Intelligence Lab (CSAIL), extracted info from 2.3 million Flickr photos to develop an algorithm that can reliably predict a photo's viewcount based on both social context and actual image content. (You can go to his research page to upload your own photo and see its relative popularity rating.)
Previous analyses had been able to predict popularity based on a user's relative social influence, but Khosla's algorithm is the first to be able to make meaning out of the complexities of actual image content - elements as specific as color, gradient, and the types of objects depicted.
Khosla uses computational techniques based on "deep learning," an emerging field that has drawn the attention of Facebook and Google for teaching computers how to learn tasks and find patterns on their own.
Among Khosla's data-driven recommendations:
- Use brighter colors (images with yellow and pink were significantly more popular than softer colors on the blue-green spectrum)
- Sex sells (among the 1000 "object classes" analyzed, the ones with the most "positive impact" included bikinis, bras and miniskirts)
- People like people (heatmaps visualizing the contributions of different regions of an image to its popularity skew towards regions that have humans and away from regions with scenery)
- Add tags (Flickr photos with more tags show up more in search and therefore get more views)
At the moment, the algorithm can't predict quite as accurately based on content alone; the fact that a photo is of a person instead of a cup will not be as strong of a predictor of the photo's popularity as, for example, the fact that it was posted by a user who has 5,000 contacts.
Khosla hopes to soon be able to automatically modify an image to make it objectively more popular, much in the same way that he and other CSAIL researchers created an algorithm that can modify your headshot to make it significantly more memorable.