A dataset of region-annotated scientific articles from PubMed Central, for document layout analysis/segmentation.
Nine document regions are annotated:
- Title
- Authors
- Abstract
- Body
- Figure
- Figure Caption
- Table
- Table Caption
- References
A script is included to download the corresponding article PDFs from PubMed Central, as well as render the article pages to JPG images. Some Linux utilities are used: curl, pdfinfo (from poppler-utils), and convert (from imagemagick). Note that by default ImageMagick disables support for PDF files, but this can easily be fixed by updating its policy file:
sudo sed -i 's/rights="none" pattern="PDF"/rights="read|write" pattern="PDF"/' /etc/ImageMagick-6/policy.xml
C.X. Soto and S.J. Yoo, "Visual Detection with Context for Document Layout Analysis", proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), 2019.