Cifar10Mnist Dataset Card

Dataset Summary

Cifar10Mnist is a synthetic image dataset created by overlaying MNIST digit images on top of CIFAR-10 images. Each example contains a 32x32 RGB image and a paired label tuple: the original CIFAR-10 class name plus the MNIST digit label.

Supported Tasks

• Image classification
• Multi-label classification
• Transfer learning
• Synthetic data research

Languages

• Not language-specific

Dataset Structure

Features

• image: RGB image of shape (32, 32, 3)
• label: tuple containing:
  • cifar_label: CIFAR-10 class name
  • mnist_label: MNIST digit label (0–9)

Split

• train: 50,000 synthetic training examples
• validation: Not explicitly split by the generator script
• test: 10,000 synthetic test examples

Dataset Construction

Curation Rationale

This dataset was created to experiment with combined visual features drawn from CIFAR-10 and MNIST. The MNIST digit is resized to 32x32 and composited over the CIFAR image using a transparency mask.

Source Data

• CIFAR-10: 32x32 RGB images in 10 object classes
• MNIST: 28x28 grayscale digit images in 10 digit classes

Annotation

Labels are derived from the original datasets:

• CIFAR-10 class label is preserved and converted to a human-readable class name (e.g. Airplane, Dog).
• MNIST digit label is preserved as an integer 0–9.

Personal or Sensitive Information

• No personal or sensitive information is present.

Dataset Uses

Recommended Uses

• Training models on multi-source synthetic image classification
• Evaluating models on low-resolution composite imagery
• Research on label conditioning and mixed-domain learning

Limitations

• Images remain low resolution (32x32) because CIFAR-10 is low-resolution.
• The dataset is synthetic and may not reflect real-world visual complexity.

Creator

• Created by the MDMTN project using the Create_Cifar10Mnist_dataset.py script.
• Repository: https://github.com/salomonhotegni/MDMTN/

Citation

If you use this dataset, please cite the dataset creator or the project repository that generated it.

@INPROCEEDINGS{10650994,
  author={Hotegni, Sedjro S. and Berkemeier, Manuel and Peitz, Sebastian},
  booktitle={2024 International Joint Conference on Neural Networks (IJCNN)}, 
  title={Multi-Objective Optimization for Sparse Deep Multi-Task Learning}, 
  year={2024},
  doi={10.1109/IJCNN60899.2024.10650994}
}

License

• Use the original licenses of CIFAR-10 and MNIST for dataset redistribution.
• If this dataset is only for internal use, follow the terms of the CIFAR-10 and MNIST datasets.