Meta-Album OmniPrint-MD-6: Synthetic Character Images for Few-Shot OCR

Meta-Album OmniPrint-MD-6 is a micro benchmark dataset containing 800 synthetic RGB images of characters across 20 categories, each sized 128x128 pixels. It was synthesized using the OmniPrint tool with specific parameters like perspective transformation and random rotation. The dataset was created by Haozhe Sun and released under the Meta-Album project in June 2021.

Use Cases

• Benchmarking few-shot image classification models based on the 20-class structure.
• Training OCR models on synthetic data with controlled nuisance parameters like rotation and shear.
• Studying the impact of synthetic data augmentation techniques on model generalization.

Strengths

• Contains 800 images with consistent 128x128 RGB resolution.
• Provides 20 distinct classes, suitable for few-shot learning tasks.
• Synthesis parameters are explicitly documented, allowing for reproducibility.

Limitations

• Dataset size is relatively small (800 images), which may limit training for larger models.
• Data is entirely synthetic, which may not fully capture the complexity of real-world character images.
• Column-level documentation is absent; field semantics must be inferred after download.

Provenance

Source

OmniPrint synthesis tool, using images from a personal mobile phone.

Collection Method

Synthesized programmatically with no further processing.