---
license: other
task_categories:
  - image-to-text
tags:
  - table-recognition
  - table-structure-recognition
  - document-ai
  - ocr
---

# Rebuilt Table Sources — OmniDocBench / CC-OCR / OCRBench-v2

1,635 table images with OTSL **and** HTML ground truth, covering the four sources
of the merged benchmark whose tables had to be **rebuilt from upstream
annotations** rather than read from a published image+html pair.

| benchmark | rows | origin |
|---|---:|---|
| `omnidocbench16` | 665 | OmniDocBench v1.6 |
| `omnidocbench15` | 512 | OmniDocBench v1.5 |
| `ccocr` | 300 | CC-OCR |
| `ocrbenchv2` | 158 | OCRBench-v2 table-parsing tasks with an explicit bbox |
| **total** | **1,635** | |

## Provenance

**OmniDocBench v1.5 / v1.6** — the releases ship page images plus a JSON whose
table entries carry a `poly` and an `html` field. Every entry with
`category_type == "table"` contributes one crop, taken with its `poly` plus a
10px margin so cell borders survive. Ground truth is the entry's own `html`;
OTSL is derived from it. Note that `opendatalab/OmniDocBench` on Hugging Face
publishes only the v1.5 JSON — v1.6 was built from the local copy.

**CC-OCR** — 300 tables, four subsets of 75: `photo_chn`, `photo_eng`,
`scan_chn`, `scan_eng`. HTML ships with the release; OTSL is derived from it.

**OCRBench-v2** — this is a general document-VQA benchmark, so most of it is
unusable for table recognition. Of its 700 `table parsing` tasks, the ones that
name an explicit region in the prompt
(`"... in the region [x1,y1,x2,y2] of the image"`, normalised to 0–1000) are
the only ones with a croppable table — **158 of them**, which is this set. The
other 542 say "all tables in the image" and were left out.

Ground truth is the `answers[0]` field. **19 of the 158 rows are Markdown
tables rather than HTML**; for those the `html` column holds the source text
verbatim and `otsl` was produced from it with a Markdown-aware converter.

## Schema

| column | type | meaning |
|---|---|---|
| `bytes` | binary | the cropped table image |
| `source` | string | original dataset name |
| `benchmark` | string | one of the four above |
| `sample_id` | string | upstream identifier |
| `rows` / `cols` | int64 | table grid dimensions |
| `otsl` | string | OTSL ground truth |
| `html` | string | HTML ground truth |
| `resolution` | string | `WIDTHxHEIGHT` |
| `width` / `height` | int64 | image dimensions |
| `structure_only` | bool | true when the table has no cell text at all |

## Scoring

`TEDS = 1 - TD / max(N_pred, N_gt)` over the table tree; TEDS-S is the
structure-only variant. When `structure_only` is true every cell is empty and
only the grid is annotated, so **TEDS is not computed** for those rows — only
TEDS-S is.

## Related

- [`CheneyWong/table_recognition_eval`](https://huggingface.co/datasets/CheneyWong/table_recognition_eval)
  — the full merged benchmark, 184,074 tables across 17 sources
- [`Cheney20/paddleocr-vl-table-recognition`](https://github.com/Cheney20/paddleocr-vl-table-recognition)
  — scoring code and GT converters

## Citation

Please cite the original datasets. This repository only repackages them under a
unified schema.