Weakly-Supervised-I2I-Translation

Code for "Extremely Weak Supervised Image-to-Image Translation for Semantic Segmentation"

Introduction

Code to train I2I translation model using both paired and unpaired training samples. This code builds on top of pytorch-CycleGAN-and-pix2pix .

Prerequisites

1. python 3.5.5
2. pytorch 0.4
3. visdom 0.1.7
4. dominate 2.3.1

Pretrained models

1. cityscapes_1
2. cityscapes_10
3. cityscapes_20

Note: Suffix indicates number of paired samples used. All models use 200 unpaired samples.

Citation

If case you use this code for your research, please cite:

@article{shukla2019extremely,
  title={Extremely Weak Supervised Image-to-Image Translation for Semantic Segmentation},
  author={Shukla, Samarth and Van Gool, Luc and Timofte, Radu},
  journal={arXiv preprint arXiv:1909.08542},
  year={2019}
}