Meta-Album OmniPrint-MD-6: 28,120 Synthetic Character Images for Few-Shot Learning

Meta-Album OmniPrint-MD-6 is a synthetic dataset of 28,120 RGB images (128x128 pixels) across 703 character categories, created using the OmniPrint synthesizer. The dataset was authored by Haozhe Sun and released in June 2021 under a CC BY 4.0 license. It is part of the Meta-Album collection for few-shot image classification.

Use Cases

• Training few-shot image classification models based on 703 distinct character categories.
• Benchmarking OCR model robustness to synthetic transformations like rotation and perspective distortion.
• Developing meta-learning algorithms using a domain-specific (OCR) image dataset with controlled nuisance parameters.

Strengths

• Contains 28,120 images, providing a substantial volume of data for model training.
• Includes 703 distinct classes, offering a high level of categorical diversity for classification tasks.
• Images are generated with precisely documented synthesis parameters (e.g., rotation within -60 to 60 degrees, specific margins), ensuring reproducibility.

Limitations

• All images are synthetically generated; model performance may not directly translate to real-world, natural character images.
• Column-level documentation is absent; the exact data structure and potential metadata fields must be inferred after download.
• The dataset's specific split into training, validation, and test sets is not detailed in the provided description.

Provenance

Source

OmniPrint synthesizer, created by Haozhe Sun.

Collection Method

Images were synthesized using the OmniPrint tool with a specific set of parameters, using foreground and background images from a personal mobile phone.