OmniDocBench
A CVPR 2025 benchmark for evaluating document-parsing systems across 1,651 pages of diverse real-world documents, with detailed block- and span-level annotations and evaluation code.
OmniDocBench is a benchmark for evaluating document parsing in real-world scenarios. It covers 1,651 PDF pages across 10 document types, 5 layout types and 5 languages, including academic papers, financial reports, newspapers, textbooks and handwritten notes.
It provides localization and recognition annotations for 28 block-level elements, such as text paragraphs, headings and tables, and 4 span-level elements, such as inline formulas and subscripts, plus reading-order and attribute-tag annotations at the page and block level. Annotation quality is checked through manual screening, intelligent and manual annotation, and expert and model review.
It ships end-to-end and single-module evaluation code covering layout detection, table recognition, formula recognition and text OCR, scored with metrics including normalized edit distance, BLEU, METEOR, TEDS and COCO detection metrics.