WayneGuo0011/TopoBench-180

TopoBench

TopoBench is a curated benchmark for topology extraction from diagram images. This release contains 162 annotated samples: 69 include redistributable images, while 93 provide annotations and source links without redistributing the images. Each annotation stores nodes and directed relations in a node-centric JSON format:

The public release contains fewer than the original 180 reviewed samples primarily because internal diagrams were excluded from public distribution.

Sample Gallery

The interactive Dataset Viewer exposes the same 162 test samples with image availability, graph annotations, source URLs, license status, and node/relation counts. Images are embedded for the 69 redistributable samples; the 93 link-only rows intentionally show an empty image cell.

{
  "Node Name": {
    "relations": [
      {
        "source": "Node Name",
        "target": "Another Node",
        "type": "directed"
      }
    ]
  }
}

Release Layout

This release intentionally separates redistributable images from link-only samples.

All 162 released samples belong to the test split. No train or validation split is included; ai2d_web and network are source domains, not dataset splits.

• datasets/ai2d_web/ and annotations/ai2d_web/: 54 AI2D-derived samples with images and annotations.
• datasets/network/ and annotations/network/: 15 network samples with recovered explicit redistribution license evidence.
• link_only/network/network_link_only_manifest.jsonl: 93 network samples where the original image is not redistributed. Each row provides image_url, source_page, provenance metadata, and the local annotation file.
• manifest.jsonl: one row per released sample, including whether the image is included or link-only.
• data/test-00000-of-00001.parquet: Viewer-ready table for the only dataset split, test.
• NETWORK_LICENSE_AUDIT.csv: provenance/license audit for all 108 network candidate samples.
• ATTRIBUTION_NETWORK_REDISTRIBUTED.csv: attribution metadata for the 15 redistributed network images.

Licensing And Attribution

AI2D is listed by the Registry of Open Data on AWS as CC BY-SA: https://registry.opendata.aws/allenai-diagrams/

The 15 redistributed network images have recovered explicit license evidence from Wikimedia Commons or the Roboflow dataset README. See ATTRIBUTION_NETWORK_REDISTRIBUTED.csv and NETWORK_LICENSE_AUDIT.csv for per-file details.

For the 93 link-only network samples, the image files are not included because redistribution rights were not verified. The dataset provides source URLs/pages and topology annotations only. Users are responsible for checking the source page terms before downloading or reusing those images.

Link-Only Image Downloader

The included standard-library script can reconstruct the link-only image directory for local research. It validates that responses are images, limits download size, retries temporary failures, and writes a JSONL log.

Preview the first three planned downloads without network access:

python scripts/download_link_only_images.py --dry-run --limit 3

Download all link-only images:

python scripts/download_link_only_images.py --acknowledge-license-risk

The acknowledgement is required because these images are not redistributed by this repository. Downloading does not grant redistribution rights; check the corresponding source_page terms before use.

Counts

Domain	Images included	Link-only	Annotations
ai2d_web	54	0	54
network	15	93	108
total	69	93	162

Quality Checks

The current release candidate was checked for:

• image/annotation pairing
• JSON parseability
• missing source, target, or type
• null relation fields
• dangling edges pointing to nonexistent nodes
• empty {} relation objects
• duplicate exact edges
• Viewer Parquet row count, test-only split, and embedded image bytes

All of the above structural checks pass for the mixed release candidate.

Citation

If you use TopoBench, please cite:

@misc{guo2026topoagentstructureawareperceptiontoreasoningframework,
      title={TopoAgent: A Structure-Aware Perception-to-Reasoning Framework for Diagram-to-Graph Topology Extraction with Large Vision-Language Models},
      author={Bangwei Guo and Xujiang Zhao and Yanchi Liu and Wei Cheng and Shengyu Chen and Dongyue Li and Masaharu Morimoto and Takayuki Kuroda and Dimitris Metaxas and Haifeng Chen},
      year={2026},
      eprint={2608.28701},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2608.28701},
}

If you use the AI2D-derived samples, also cite AI2D and follow its CC BY-SA attribution/share-alike requirements.