gansulishuzhi

A small out-of-distribution benchmark for vision-language model figure-bbox detection. The input is a 200 DPI scan of a 1970s Chinese botanical book page; the model must output pixel-coordinate bounding boxes for every figure on the page (photos, line drawings, illustrations). No OCR is evaluated.

• 61 pages, 64 figures, all bboxes hand-verified
• Page resolution: 1395×2037 px (200 DPI render)
• Language: Chinese (scan content; English / Chinese prompt examples below)

Why OOD

Modern VLM grounding stacks train on clean web documents. These pages have:

• 50-year-old print, ink bleed, scanner noise
• Mixed figure styles: black-and-white photos, line-drawn botanical illustrations, small inset diagrams
• Handwritten margin notes
• Irregular layout — captions and body text interleaved, page numbers, section headings

It's a narrow but unforgiving probe of whether a VLM's grounding generalises beyond clean modern documents.

Files

manifest.jsonl     61 rows, one per page
pages/             page_<NNN>.png   — model input
overlays/          page_<NNN>.png   — same image with red GT bboxes (visual QA)

Manifest schema

manifest.jsonl, one JSON object per line:

{
  "id": "page_014",
  "page": 14,
  "image_path": "pages/page_014.png",
  "image_size_px": [1395, 2037],
  "gt_bboxes": [[168, 264, 719, 1055], [456, 1199, 1247, 1751]]
}

gt_bboxes is [[x0, y0, x1, y1], ...] in image-pixel coords, origin top-left, (x0,y0) = top-left, (x1,y1) = bottom-right.

How to load

from huggingface_hub import snapshot_download
import json
from pathlib import Path

root = Path(snapshot_download(repo_id="yunfengwang/gansulishuzhi", repo_type="dataset"))
samples = [json.loads(l) for l in (root / "manifest.jsonl").read_text().splitlines() if l.strip()]
print(samples[0])
# image at: root / samples[0]["image_path"]

Suggested prompt

Ask the model for a JSON list of [x0, y0, x1, y1] integer pixel coords at the actual image resolution. A working Chinese prompt:

这是一页扫描书页，可能包含 0 张、1 张或多张图（照片、线描、插图等，不含正文/标题/页眉/页脚/表格）。 请检测页面中每一张图的边界框，以 JSON 列表输出，每个元素为 [x0, y0, x1, y1] 整数像素坐标。 坐标原点在图像左上角，x 向右、y 向下；相对当前图像分辨率 (W×H)。仅框住图本身，不要包含 caption 文字或周围正文。若页面无图，输出 []。 只输出 JSON，不要任何解释或代码块标记。

Recommended metric

Greedy max-IoU match between predicted and ground-truth bboxes. For threshold T: a matched pair counts as TP iff its IoU ≥ T. Report micro precision / recall / F1 at IoU 0.3 / 0.5 / 0.7, plus mean IoU of matched pairs. No mAP because the suggested output schema has no per-bbox confidence.

A reference implementation lives in the companion code repo: github.com/vra/gansulishuzhi (uv sync && uv run eval.py --model-id <id> --output runs/<name>.jsonl).

License

The page images are scans of a 1970s Chinese reference work on Gansu pear cultivars. They are provided for non-commercial research use only. Downstream users are responsible for compliance with applicable copyright law in their jurisdiction. The accompanying manifest, code, and metrics are released under the MIT license.

Citation

If you use this dataset, please cite the companion repo:

@misc{gansulishuzhi2026,
  title  = {gansulishuzhi: An OOD benchmark for VLM figure-bbox detection on 1970s scanned Chinese book pages},
  author = {Wang, Yunfeng},
  year   = {2026},
  url    = {https://huggingface.co/datasets/yunfengwang/gansulishuzhi}
}