Evaluating Temporal Information for Social Image Annotation and Retrieval

Can we use the temporal evolution of annotations in Web images to improve tasks such as annotation, indexing and retrieval? This important question is the main motivation for this work. Typically visual content, text and metadata, are used to improve these tasks. A characteristic that has received less attention, so far, is the temporal aspect of social media production and tagging. The main contribution of this paper is a thorough analysis of the temporal aspects of two popular datasets commonly used for tasks such as tag ranking, tag suggestion and tag refinement, namely NUS-WIDE and MIR-Flickr-1M. The correlation of the time series of the tags with Google searches shows that for certain concepts web information sources may be beneficial to annotate social media.