Understanding image virality

This work trains classifiers with state-of-the-art image features to predict virality of individual images, relative virality in pairs of images, and the dominant topic of a viral image, and compares machine performance to human performance on these tasks.

Abstract

Virality of online content on social networking websites is an important but\nesoteric phenomenon often studied in fields like marketing, psychology and data\nmining. In this paper we study viral images from a computer vision perspective.\nWe introduce three new image datasets from Reddit, and define a virality score\nusing Reddit metadata. We train classifiers with state-of-the-art image\nfeatures to predict virality of individual images, relative virality in pairs\nof images, and the dominant topic of a viral image. We also compare machine\nperformance to human performance on these tasks. We find that computers perform\npoorly with low level features, and high level information is critical for\npredicting virality. We encode semantic information through relative\nattributes. We identify the 5 key visual attributes that correlate with\nvirality. We create an attribute-based characterization of images that can\npredict relative virality with 68.10% accuracy (SVM+Deep Relative Attributes)\n-- better than humans at 60.12%. Finally, we study how human prediction of\nimage virality varies with different `contexts' in which the images are viewed,\nsuch as the influence of neighbouring images, images recently viewed, as well\nas the image title or caption. This work is a first step in understanding the\ncomplex but important phenomenon of image virality. Our datasets and\nannotations will be made publicly available.\n