Component-Level OBI Retrieval

Source code and dataset for ICMR'24 paper "Component-Level Oracle Bone Inscription Retrieval" (Best Paper Candidate)

You may find our other works on OBI:

• [AAAI'25] Component-Level Segmentation for Oracle Bone Inscription Decipherment

0. To Do List

1. Task Definition
2. dataset
3. code

1. Task Definition

Oracle Bone Inscriptions (OBIs) are pictographic scripts, various realworld entities are abstracted into distinct components within OBIs to represent different meanings, such as "foot", "person", and "water". Subsequently, a more nuanced semantics is conveyed through the combination of these components. Figure below illustrates four examples, i.e., "stand", "pace", "wade", and "climb", of deciphered OBIs.

• 企 Stand is constituted by one "止 foot" component and one "人 person" component, symbolizing an individual standing upright.
• 步 Pace is formed by two "止 foot" components arranged in a front-and-back configuration, symbolizing an act of stepping once.
• 涉 Wade is composed of two "止 foot" components in a front-and-back arrangement, with one "水 water" component intersecting through them, symbolizing walking into the water.
• 陟 Climb is constructed with two upward-facing "止 foot" components and one "阜 mound" component, symbolizing ascending a mound.

Each of them is composed of the "foot" component along with other distinct components. This diversity in constituent components results in distinct meanings for these characters. However, since these characters all incorporate the "foot" component, they share a certain semantic association with each other. Therefore, linking different OBIs through their respective components holds significant implications. Specifically, exploring the patterns among characters sharing the same component facilitates the deciphering and comprehension of OBIs.

To this end, we propose Component-level Oracle Bone Inscription Retrieval, i.e., given a component, the goal is to retrieve all OBIs that incorporate this particular component.

2. Dataset

2.1. Introduction

In OBI Component 20, we have selected 20 common OBI components. Due to the different forms each component can take, we chose representative forms to display in the following diagram.

Then, we collected 11,335 OBI character images from the 小學堂 based on these components. We invited Prof. Pui-ling Tang and Ms. Peiying Zhang from the School of Chinese, the University of Hong Kong to screen these characters, removing images that did not contain the 20 selected components, leaving us with 9,245 OBI character images. Within these images, Ms. Zhang further annotated the specific positions of the components in 1,012 OBI character images, striving to cover the different forms of the same component. Ultimately, OBI Component 20 contains a total of 10,257 OBI images, of which 9,245 are OBI characters and 1,012 are OBI components. Their distribution is as shown in the table below.

ID	Component	# Character	# Component	ID	Component	# Character	# Component
0	日	371	18	10	女	641	29
1	月	106	41	11	子	179	33
2	雨	152	17	12	目	422	31
3	阜	115	16	13	攴	414	91
4	水	622	41	14	止	1132	72
5	屮	267	14	15	衣	69	51
6	木	465	24	16	口	1592	42
7	犬	204	117	17	王	55	8
8	大	385	32	18	矢	383	32
9	人	1403	226	19	刀	268	77

For more details about the dataset, please refer to here.

2.2. Request for Dataset

Considering copyright issues, if you need to use this dataset, please provide the following information (either in Chinese or English) in an email to [email protected] (Dr. Zhikai Hu, 胡志锴) AND c.c. [email protected] (Ms. Peiying Zhang, 張沛瑩). We will provide you with the dataset download link within 5 working days after receiving your email (It should be a valid .edu email that matches your institution):

1. your name,
2. your institution,
3. the intended use of the dataset, and
4. a declaration ensuring that it will not be used for commercial profit.

3. Code

Train the model:

python train.py --componet_path components_file_name /
                --character_path characters_file_name /
                --epoch 80 /
                --batch_size=32 /
                --num_class=20

Test the model:

python test.py --componet_path components_file_name /
               --character_path characters_file_name

Visualize the retrieval results:

python visual.py --componet_path components_file_name /
                 --character_path characters_file_name /
                 --k=10

4. Citation

@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.