SCI-3000: A Dataset for Figure, Table and Caption Extraction from Scientific PDFs

Abstract

Extracting figures and similar visual elements from PDFs of scientific publications is important but non-trivial, and progress is impeded by a lack of datasets for evaluation and machine learning. In this work, we describe and publish the SCI-3000 dataset, containing 3 000 PDFs of scientific publications (34 791 pages) with annotations of figures, tables, and corresponding captions, from the fields of computer science, biomedicine, chemistry, physics, and technology. We demonstrate the use of the dataset to benchmark two figure, table, and caption extraction approaches from recent literature: one rule-based and one deep learning-based.