Meta-Album OmniPrint-MD-mix: 28,240 Synthetic Character Images for Few-Shot Learning

OmniPrint-MD-mix is a synthetic dataset of 28,240 RGB images (128x128 pixels) across 706 character categories, designed for few-shot learning benchmarks. Created by Haozhe Sun using the OmniPrint synthesizer in June 2021, it applies a specific set of nuisance parameters like perspective transformation, rotation, and color variations to images of photos taken by a personal mobile phone. The dataset is released under a CC BY 4.0 license as part of the Meta-Album multi-domain collection.

Use Cases

• Benchmarking few-shot image classification models based on the 706 distinct character categories.
• Training models for OCR robustness based on synthesized nuisance parameters like rotation and perspective transformation.
• Developing domain adaptation techniques for synthetic-to-real transfer based on the OmniPrint generation method.
• Researching meta-learning algorithms using the structured 20-settings-per-class synthesis approach mentioned in the description.

Strengths

• Contains 28,240 images, providing a substantial volume for model training and evaluation.
• Covers 706 distinct character classes, offering high categorical diversity for classification tasks.
• Images are synthesized with a controlled, documented set of parameters (e.g., rotation within -60 to 60 degrees, specific margins), allowing for reproducible augmentation studies.
• Released under the permissive CC BY 4.0 license, facilitating open sharing and derivative use.

Limitations

• All images are synthetic, generated from photos taken by a personal mobile phone, which may not fully represent real-world OCR data distributions.
• Column-level documentation is absent; field semantics for any potential metadata must be inferred after download.
• Last update date is unknown; freshness unverified.

Provenance

Source

OmniPrint synthesizer, using photos taken by a personal mobile phone.

Collection Method

Synthetic generation via the OmniPrint tool with specified parameters for perspective, rotation, shear, and color transformations.

Time Range

Created on 25 June 2021.