GoBo Synthetic Handwriting Datasets Generated by Style Transfer

This record contains synthetic handwritten word image datasets generated by applying handwriting style transfer to writers from the GoBo dataset (https://zenodo.org/records/8085511).

The purpose of the dataset is to extend the available handwritten data for each GoBo writer by generating additional synthetic word images in the corresponding writer style. 

The record contains three synthetic handwriting datasets:

01_GoBo_synthetic_one_shot.zip  

Synthetic handwriting generated with the One-DM one-shot handwriting generation model (https://github.com/dailenson/One-DM). For each writer, one randomly selected GoBo style reference image was used. The vocabulary is based on the original GoBo word groups: brown, cedar, domain_A_train, domain_B_test, and nonwords.

02_GoBo_synthetic_few_shot.zip  

Synthetic handwriting generated with the VATr few-shot handwriting generation model (https://github.com/aimagelab/VATr). For each writer, 15 randomly selected GoBo style reference images were used. The vocabulary is based on the original GoBo word groups: brown, cedar, domain_A_train, domain_B_test, and nonwords.

3_GoBo_synthetic_few_shot_10k_common_words.zip  

A larger few-shot synthetic handwriting dataset generated with VATr. For each writer, 15 randomly selected GoBo style reference images were used. The vocabulary contains 10,000 common English words from the Leipzig Wortschatz project (https://wortschatz.uni-leipzig.de/en/download/English). This archive contains 40 writers × 10,000 images = 400,000 PNG images.

The record also includes the file AttentionHTR_IAM_baseline_model.pth. This file contains the generic AttentionHTR baseline model trained on the IAM dataset and used as the starting point for the personalization experiments. The baseline model was not used to generate the synthetic images.