This page contains models and datasets to reproduce our experiments presented in the paper Synthesis in Style: Semantic Segmentation of Historical Documents using Synthetic Data.
The code for the paper can be found here. ​

The file benchmark_dataset.zip contains the dataset that was used in our experiments. The original and ground truth images can be found in the respective subdirectories.
The files in-domain-images.txt and out-of-domain-images.txt indicate which of these images belong to the in-domain and out-of-domain split of our dataset (see paper for more information). ​

Besides the dataset, we provide all of our trained models. The first part of the filename indicates the synthesis method, either sis (Synthesis in Style/ours) or datasetgan. The second part of the filename refers to the underlying GAN model.

In our paper, we only evaluated StyleGAN, but we also synthesized data using SWAGAN.
All archives contain the checkpoints for the three segmentation models: DocUFCN, EMANet, and TransUNet.