Oracle-MNIST

Overview

The Oracle-MNIST dataset contains ancient Chinese oracle-bone characters, about 3,000 years old, from 10 categories. Heavy noise and distortion from burial and aging, along with highly varied writing styles, make classification harder than MNIST. Like MNIST, it uses gzipped IDX files containing 28x28 grayscale uint8 images and integer labels.

• Train: 27,222 images

• Test: 3,000 images (300 per class)

Data Structure

When accessing an example using ds[i], you will receive a dictionary with the following keys:

Key	Type	Description
image	PIL.Image.Image	28x28 grayscale image with uint8 pixel values
label	int	Character class label (0-9)

Character Classes

Class names follow the label order in Table 2 of the paper:

Label	Class Name	Train	Test
0	big	2,433	300
1	sun	2,765	300
2	moon	2,668	300
3	cattle	2,614	300
4	next	2,610	300
5	field	2,328	300
6	not	2,710	300
7	arrow	2,360	300
8	time	3,335	300
9	wood	3,399	300

Usage Example

from stable_datasets.images import OracleMNIST

# First run will download and prepare the dataset cache
train_ds = OracleMNIST(split="train")
test_ds = OracleMNIST(split="test")

sample = train_ds[0]
print(sample.keys())  # {"image", "label"}

class_names = train_ds.info.features["label"].names
print(class_names[sample["label"]])

# Optional: make it PyTorch-friendly
train_torch = train_ds.with_format("torch")

Related Datasets

• MNIST: Original handwritten digit dataset with the same format and dimensions

• KMNIST: Kuzushiji-MNIST, a drop-in replacement with cursive Japanese characters

• Fashion-MNIST: Drop-in replacement with clothing images

References

• Homepage: https://github.com/wm-bupt/oracle-mnist

• License: Apache-2.0

• Downloads are pinned to commit b6cb263f84d39a8bbb94c5b806d846b600068924 and verified with SHA-256 checksums.

Citation

@article{wang2024dataset,
  title={A dataset of oracle characters for benchmarking machine learning algorithms},
  author={Wang, Mei and Deng, Weihong},
  journal={Scientific Data},
  volume={11},
  number={1},
  pages={87},
  year={2024}
}