OmniPrint-MD-5-bis: Synthetic Character Images for Few-Shot Learning

800 synthetic RGB images of characters across 20 categories, each 128x128 pixels. The dataset was created using the OmniPrint synthesizer with specific nuisance parameters like random perspective, translation, and rotation, and released by Haozhe Sun in June 2021. It is part of the Meta-Album benchmark for few-shot image classification.

Use Cases

• Benchmarking few-shot image classification models based on the described 20-class, 800-image structure.
• Training models for character recognition robust to synthetic nuisances like rotation and shear mentioned in the description.
• Developing domain adaptation techniques for OCR using data with Poisson Image Editing blending and random color foregrounds.
• Studying the impact of synthetic data generation parameters on model performance in controlled settings.

Strengths

• Image count and class count are explicitly stated (800 images, 20 classes).
• Image dimensions and format are clearly defined (128x128 RGB).
• Synthesis parameters are detailed, providing reproducibility for the data generation process.
• Released under a permissive CC BY 4.0 license.

Limitations

• Dataset consists entirely of synthetic images, which may not fully capture the complexity of real-world OCR data.
• Column-level documentation is absent; field semantics must be inferred after download.
• Last update date is unknown; freshness unverified.

Provenance

Source

OmniPrint synthesizer (https://github.com/SunHaozhe/OmniPrint)

Collection Method

Synthesized programmatically with specified parameters; backgrounds are images downloaded from Pexels.