Scientific poster generation: A new dataset and approach

Abstract

Automating poster creation from research papers saves scientists time. However, training models for this task is challenging due to limited datasets. Moreover, existing methods are mostly rule/template-based, which lack the flexibility to adapt to different content and design requirements in scientific posters. Our contributions aim to address these issues. We introduce Sci-PosterLayout, a dataset comprising 1,226 scientific posters with greater variety in content, layout and domains. Using a template-free method with a seq2seq model and Design Pattern Schema (DPS), we learn various content and design patterns for poster layout generation. Evaluations against existing methods and datasets show our approach produces high-quality posters with diverse layouts. Our work seeks to advance research in scientific poster generation by building a new dataset and proposing template-free methods that require minimal human intervention. The Sci-PosterLayout dataset will be publicly available at https://github.com/kitman0000/Sci-PosterLayout-Data.