Meta-Album OmniPrint-MD-5-bis: 28,240 Synthetic Character Images

OmniPrint-MD-5-bis is a dataset of 28,240 synthetic RGB images (128x128 pixels) across 706 character categories, generated using the OmniPrint synthesizer. It was created by Haozhe Sun in June 2021 as part of the Meta-Album collection for few-shot learning. The images are synthesized with specific nuisance parameters like perspective transformation, rotation, and random backgrounds from Pexels.

Use Cases

• Training few-shot image classifiers based on 706 distinct character categories.
• Benchmarking OCR model robustness to synthetic transformations like perspective, shear, and rotation mentioned in the description.
• Developing data augmentation strategies by studying the effects of controlled synthesis parameters like Poisson Image Editing and elastic transformation.

Strengths

• Contains 28,240 images, providing a substantial volume of data for model training.
• Includes 706 distinct character classes, enabling fine-grained classification tasks.
• Images are generated with a documented, controlled set of synthesis parameters, allowing for reproducible experiments.

Limitations

• Data is entirely synthetic, which may limit generalizability to real-world, natural character images.
• Column-level documentation is absent; field semantics must be inferred after download.
• Last update date is unknown; freshness unverified.

Provenance

Source

OmniPrint synthesizer, with backgrounds from Pexels.

Collection Method

Synthetic generation using the OmniPrint tool with specified parameters.