How to Make Your Photo Uploads More Popular: Take Your Clothes Off and Brandish a Firearm
The research, carried out by Aditya Khosla of MIT’s CSAIL with friends from Ebay and DigitalGlobe, uses computer vision and machine learning to answer the age old question “What makes an image popular?” The researchers began with a massive corpus of 2.3 million images from Flickr, and then used deep learning (neural networks) and analysis of social cues to discover what exact combination of image features and social context makes an image likable. [Research paper, PDF]
First the software looks at high-level features, namely objects. Modern computer vision algorithms are very good at picking out and identifying objects in images. Unsurprisingly, objects such as miniskirts, bras, and guns have a strong positive impact on an image’s popularity. Most cuddly animals score a “medium positive impact” rating, while wild boars, dull-looking items of food, and other everyday objects create a “low positive impact.” Spatulas, plungers, laptops, and golfcarts actually create a negative impact.