coo-comic-onomatopoeia-safetensors

COO Comic Onomatopoeia — Safetensors conversions

Unofficial tensor-only Safetensors conversions of the pretrained checkpoints published with COO: Comic Onomatopoeia Dataset for Recognizing Arbitrary or Truncated Texts (ECCV 2022).

The original authors are Jeonghun Baek, Yusuke Matsui, and Kiyoharu Aizawa. This repository is maintained by mayocream; it is not an official release by the paper authors.

Included checkpoints

The larger leaderboard in the upstream README contains other third-party models. Those unrelated checkpoints are intentionally not mirrored here.

Conversion guarantees

• Source checkpoints were downloaded from the authors' published Dropbox links.
• PyTorch files were loaded with weights_only=True.
• Only the model tensor mapping was exported; optimizer, scheduler, trainer, and iteration state were omitted.
• Tensor names, shapes, dtypes, and values were preserved exactly.
• Every output was reloaded with Safetensors and compared tensor-by-tensor using torch.equal.
• Source and converted SHA-256 hashes are recorded in conversion_manifest.json and embedded in each Safetensors header.

The conversion script is available at scripts/convert_checkpoints.py.

Loading

Safetensors contains state dictionaries, not standalone model implementations. Instantiate the matching model from the upstream repository, then load the corresponding file:

from safetensors.torch import load_file

state_dict = load_file("trba-rot-sar-hardroi-2d/model.safetensors")
model.load_state_dict(state_dict)

ABCNetv2 and MTSv3 original checkpoints store their weights under the model key. M4C-COO also stores weights under model. TRBA checkpoints are state dictionaries at the checkpoint root. These layouts are normalized here so each .safetensors file is directly a state dictionary.

Dataset and architecture notes

The models were trained or fine-tuned on COO annotations over licensed Manga109 images. Manga109 images are not included here. Detection, recognition, and link prediction were evaluated as separate tasks rather than one jointly trained end-to-end pipeline.

M4C-COO requires additional inputs from the upstream project, including fastText, PHOC, bounding-box information, and visual features extracted using MTSv3. These supporting resources are not included in this repository.

License and attribution

The upstream repository does not state a single explicit license for the pretrained weight files. Its component codebases have different terms:

• ABCNetv2 / AdelaiDet: BSD-2-Clause for academic use; upstream requests contact for commercial use.
• Mask TextSpotter v3: CC BY-NC 4.0.
• TRBA modifications: MIT; underlying deep-text-recognition-benchmark code: Apache-2.0.
• M4C-COO / MMF: BSD.
• COO annotations: CC BY 4.0.
• Manga109 images: separate Manga109 terms; images are not redistributed here.

This repository therefore uses license: other and does not grant additional rights. Users are responsible for reviewing the upstream licenses and Manga109 terms for their use case.

Citation

@inproceedings{baek2022COO,
  title={COO: Comic Onomatopoeia Dataset for Recognizing Arbitrary or Truncated Texts},
  author={Baek, Jeonghun and Matsui, Yusuke and Aizawa, Kiyoharu},
  booktitle={Proceedings of the European Conference on Computer Vision (ECCV)},
  year={2022}
}