Collecting, Analyzing and Predicting Socially-Driven Image Interestingness

Interestingness has recently become an emerging concept for visual content assessment. However, understanding and predicting image interestingness remains challenging as its judgment is highly subjective and usually context-dependent. In addition, existing datasets are quite small for in-depth analysis. To push forward research in this topic, a large-scale interestingness dataset (images and their associated metadata) is described in this paper and released for public use. We then propose computational models based on deep learning to predict image interestingness. We show that exploiting relevant contextual information derived from social metadata could greatly improve the prediction results. Finally we discuss some key findings and potential research directions for this emerging topic.