BUDDI Table Factory: A toolbox for generating synthetic documents with annotated tables and cells

Tables are the most convenient way to represent structured information in a document. Understanding the table structure is critical to understanding its contents. Several deep learning-based approaches from the literature have shown promising results in understanding table structures, but they require large amounts of annotated data. However, the availability of annotated datasets to train these methods are expensive, laborious, and very limited. Moreover, human-annotated data suffers from inconsistencies in table and cell annotations. We propose BUDDI Table Factory (BTF) for synthetically generating annotated documents with a wide range of variations in table structures. We propose a heuristics-based method to generate a variety of table structures from which we generate synthetic documents using LaTeX. We propose a computer vision-based approach to localize table and cell regions and automatically generate annotations in PASCAL VOC challenge format. We empirically illustrate the advantage of adding synthetic BTF documents with limited original documents to the model training, which can significantly improve the TEDS and IoU performance of the table structure recognition tasks in public and real-world healthcare datasets.