Multimodal visualization based on latent topic analysis

Image collection visualization is an important component of exploration-based image retrieval systems. In this paper we address the problem of generating an image collection visualization in which images and text can be projected together. Given a collection of images with attached text annotations, our aim is to find a common representation for both information sources to model latent correlations among the collection. Using the proposed latent representation, an image collection visualization is built, in which both data modalities (images and text) can be projected simultaneously. The resulting collection visualization allows to identify the relationships between images and text terms, enabling a better understanding of the collection semantic structure. The resulting visualization scheme can be used as the core metaphor in an interactive image exploration system.