Component-Level OBI Segmentation

Source code for AAAI'25 paper "Component-Level Segmentation for Oracle Bone Inscription Decipherment"

0. To Do List

1. Task Definition
2. Code
3. Dataset

1. Task Definition

Background

Prof. Huang Tianshu pointed out that OBIs that exhibit direct correspondences with modern Chinese characters were basically deciphered by the late 20th century. The remaining undeciphered OBIs are exceedingly difficult to decode SOLELY through glyph analysis.

Take the case where Dr. Jiang Yubin successfully diciphered the OBI (denoted as A) for example, the deciphering process roughly includes three steps:

• Step 1: identifying OBIs with similar glyph to A to make an initial semantic hypothesis;
• Step 2: seeking evidence from other OBIs to refine and support this hypothesis;
• Step 3: crossreferencing pre-Qin literature to find corresponding corpus that can further validate the hypothesis from the Step 2.

The final deciphered meaning 蠢（蠢动，骚动） differs significantly from A in terms of glyph structure, indicating that accurate interpretation cannot be achieved through glyph evolution alone.

Why this task?

One of the key contributions of Dr. Jiang Yubin lies in revising the erroneous component segmentation of OBI B from previous study during this step, thereby enabling the textual evidence in Step 3 to perfectly align with the corrected segmentation, as shown in the above figure. Thus, accurately segmenting OBIs to extract the target components is crucial. We termed this task Component-Lvel OBI Segmentation.

2. Code

Preparation

Installing the MMlab as follow:

pip install mmcv-full==1.4.7 -f https://download.openmmlab.com/mmcv/dist/cu113/torch1.11.0/index.html
pip install mmsegmentation==0.24.0

You can download other mmlab version for your cuda and torch via

https://download.openmmlab.com/mmcv/dist/{your_cuda_version}/{your_torch_version}/index.html

Then, you need download pvt-medium as backbone.

Train

python train.py --data_path your_data_path

Test

python test.py --data_path your_data_path

Visualization

python visual.py --data_path your_data_path

3. Dataset

Please refer to Component-Level Oracle Bone Inscription Retrieval to apply for the dataset and here for more details about the dataset.

4. Citation

@inproceedings{hu2025component,
  title={Component-Level Segmentation for Oracle Bone Inscription Decipherment},
  author={Hu, Zhikai and Cheung, Yiu-ming and Zhang, Yonggang and Zhang, Peiying and Tang, Pui-ling},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={},
  number={},
  pages={},
  year={2025}
}
@inproceedings{hu2024component,
  title={Component-Level Oracle Bone Inscription Retrieval},
  author={Hu, Zhikai and Cheung, Yiu-ming and Zhang, Yonggang and Zhang, Peiying and Tang, Pui-ling},
  booktitle={Proceedings of the 2024 International Conference on Multimedia Retrieval},
  pages={647--656},
  year={2024}
}

Acknowledgments

We would like to thank 小學堂 for sharing the public OBI data. We are also grateful to Mr. Changxing Li for his assistance with the data collection and code implementation.