Modeling Image Appeal Based on Crowd Preferences for Automated Person-Centric Collage Creation

Abstract

This paper attempts to model Image Appeal of photos portraying people, with the objective of automatically ranking and selecting the most appealing ones for the creation of interesting person-centric collages/collections. To understand the notion of image appeal, we employed crowdsourcing, using 350 workers who were asked to select a representative subset of images from five different person-centric album themes (involving a man, woman, couple, girl and baby). The albums were previously balanced with respect to nine different image attributes using Binary Integer Programming. The crowdsourcing study revealed identifiable patterns in the photo selection process, with more appealing photos securing more hits than less appealing ones. We then employed nine low-level image features and Support Vector Regressors to model photo selection statistics-the best model explained 63% of the selection patterns, and our analyses also confirmed the role of context in influencing Image Appeal. Finally, Image Appeal predictions on unseen photos are presented to demonstrate the promise of our approach.