gt-free-ocr-metrics/omnidocbench-render-compare

OmniDocBench Render-and-Compare

This dataset contains the rendered HTML reconstructions and comparison images produced by a render-and-compare pipeline — a reference-free visual similarity evaluation framework for OCR systems.

Overview

The pipeline processes each page of OmniDocBench through a Qwen3.5-122B-A10B OCR model, renders the structured output back to a PNG via HTML (reconstructed.png), and compares it against the original page scan (masked_original.png) using reference-free visual metrics.

Five OCR extraction variants are provided, each targeting a different subset of document element types:

Variant	Recognised elements	Pages
ocr_all	text + formula + table (with image regions masking)	1 355
ocr_all_no_mask	text + formula + table (no image masking)	1 355
ocr_text	text only	1 349
ocr_formula	formula only	200
ocr_table	table only	351

Dataset Structure

<variant>/
  <page_id>/
    masked_original.png     # original scan with non-target regions grayed out
    reconstructed.png       # HTML render of the OCR output (screenshot at 1× DPI)
    ocr_html.html           # raw HTML produced by the OCR model
    ocr_elements.json       # extracted text bboxes + content  (text variants only)
    ocr_formula_elements.json  # extracted formula bboxes + LaTeX (formula variants)
    ocr_table_elements.json    # extracted table bboxes + HTML  (table variants)

docsim_triplets/
  manifest.jsonl            # 20 280 training triplets for the DocSim LoRA head

OmniDocBench.json           # original OmniDocBench ground-truth annotation file (GT text/formula/table)

<page_id> is the identifier used in OmniDocBench.json (e.g. book_en_5.Advanced.Modern.Algebra_page_572).

Not all files are present in every variant:

• ocr_all_no_mask pages contain only masked_original.png, ocr_html.html, and reconstructed.png (no element JSON files, no bbox visualization).
• Formula/table element JSON files are absent when no such elements were detected on a page.

DocSim Triplets

docsim_triplets/manifest.jsonl is a JSONL file with 20 280 training triplets used to fine-tune a LoRA-adapted CLIP+DINOv2 similarity head (DocSim). Each line is a JSON object:

{
  "anchor_path":   "ocr_all_no_mask/<page_id>/masked_original.png",
  "positive_path": "ocr_all_no_mask/<page_id>/reconstructed.png",
  "negative_path": "ocr_all/<other_page_id>/reconstructed.png",
  "anchor_ed":   0.0,
  "positive_ed": 0.12,
  "negative_ed": 0.64,
  ...
}

Paths are relative to the root of this repository. anchor_ed / positive_ed / negative_ed are Hungarian-matched edit distances (lower = higher text fidelity) used as supervision during training.

Intended Use

• Research on reference-free OCR evaluation: masked_original / reconstructed pairs provide ground material for developing visual similarity metrics that do not require OCR ground truth.
• Document visual quality research: diverse real-world layouts (books, exams, slides, financial reports, scientific papers) in EN and ZH.
• Training document similarity models: the DocSim triplet manifest can be used to train or evaluate learned perceptual similarity heads for document images.

Out-of-Scope Use

• The reconstructed images and OCR outputs should not be used to extract or reproduce copyrighted text for redistribution.
• The dataset is not a ground-truth OCR corpus; the reconstructed.png images are model outputs, not verified transcripts.

Limitations and Biases

• Language distribution: OmniDocBench contains English and Chinese pages; the ratio of each language is determined by the original benchmark.
• Element sparsity: formula and table variants cover only pages with relevant ground-truth content; pages with no formulas or tables are absent from those variants.
• Render fidelity: HTML-to-PNG rendering uses a headless browser at a fixed viewport; fonts, rendering artefacts, and DPI settings may differ from the original scans.

Sensitive and Personal Information

OmniDocBench draws from public academic and professional documents. No personally identifiable information (PII) was deliberately collected. Financial report pages may incidentally reference company names or executives. No medical, legal, or biometric data is present.

Source Dataset

This dataset is derived from OmniDocBench (OpenDataLab / SJTU, CC-BY-NC-4.0). The OmniDocBench.json annotation file is reproduced here for reviewer convenience; original page scans are available in the source dataset.

Citation

If you use this dataset, please cite:

@article{omnidocbench2024,
  title   = {OmniDocBench: Benchmarking Document Parsing with Diverse Layouts on Real-World Data},
  author  = {Hu, Linke and others},
  journal = {arXiv},
  year    = {2024},
  eprint  = {2412.07626}
}

License

Released under CC-BY-NC-4.0. The original OmniDocBench page images (images/) retain their source licensing terms; users must comply with those when using the raw scans.