Learning to Generate Posters of Scientific Papers

Researchers summarize and represent their paper content with scientific posters, which efficiently convey their ideas. Generating a good scientific poster, however, is challenging for novel researchers, since it needs to be readable, informative, and aesthetic. This paper for the first time studies the challenging problem of learning to generate posters from scientific papers. To this end, a data-driven framework is proposed by utilizing probabilistic graphical models. Specifically, given contents to display, the key elements of a good poster, including panel layout and attributes of each panel, are learned and inferred from data. Then composition of graphical elements within each panel is synthesized. To validate our framework, we contribute a Poster-Paper dataset with exhaustively labelled attributes of poster panels. Qualitative and quantitative results indicate the effectiveness of our framework.