This is the code related to "Global and Local Texture Randomization for Synthetic-to-Real Semantic Segmentation" (IEEE TIP 2021).

Global and Local Texture Randomization for Synthetic-to-Real Semantic Segmentation
IEEE Transactions on Image Processing (TIP 2021)

If you find it helpful to your research, please cite as follows:

@article{peng2021global,
  title={Global and Local Texture Randomization for Synthetic-to-Real Semantic Segmentation},
  author={Peng, Duo and Lei, Yinjie and Liu, Lingqiao and Zhang, Pingping and Liua, Jun},
  journal={IEEE Transactions on Image Processing},
  year={2021},
  publisher={IEEE}
}

• PyTorch 1.7.1
• CUDA 10.2
• Python 3.7
• Torchvision 0.9.0
• Download Painter by Numbers which are paintings for GTR.
• You should transfer the raw source dataset into multiple-style datasets using the pre-trained style transfer network AdaIN and put the correct paths in line 685 of the python file (./tools/TR_BR.py )
• Download the model pretrained on ImageNet. Put it into each file named as (pretianed_model).

• Download GTA5 datasets, in the experiments, we crop GTA5 images to 640X640.
• Download SYNTHIA. We crop images to 640X640.
• Download Cityscapes. We resize Cityscapes images to 1024x512.
• Download BDDS. We resize BDDS images to 1024x512.
• Download Mapillary. We resize Mapillary images to 1024x512.

Open the terminal and type the following command to pretrain the model on the source domain (GTA5).

python3 tools/TR_BR.py

We present several qualitative results reported in our paper.