1,623 handwritten character classes across 50 different alphabets, each containing 20 unique samples. The data is provided as 105x105 grayscale images and is divided into background and evaluation sets for few-shot learning tasks.
Use Cases
- Train one-shot learning models using the image and label columns to recognize characters from a single sample
- Benchmark Siamese networks for image similarity using the alphabet and character_class hierarchy
- Evaluate meta-learning algorithms on their ability to generalize to new writing systems using the evaluation split
Strengths
- 1,623 unique character classes across 50 distinct alphabets
- 20 handwritten samples per character provided as 105x105 pixel grayscale images
- Organized into 'background' and 'evaluation' sets containing 30 and 20 alphabets respectively