Global and Local Texture Randomization for Synthetic-to-Real Semantic Segmentation
This is the code related to "Global and Local Texture Randomization for Synthetic-to-Real Semantic Segmentation" (IEEE TIP 2021).
Paper
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}
}
Preparation
- 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).
Datasets
- 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.
Usage
Open the terminal and type the following command to pretrain the model on the source domain (GTA5).
python3 tools/TR_BR.py
Results
We present several qualitative results reported in our paper.
