SM-PPM

This is a Pytorch implementation of our paper "Style Mixing and Patchwise Prototypical Matching for One-Shot Unsupervised Domain Adaptive Semantic Segmentation".

AAAI2022 [arxiv]

The paper does not available currently. We are working on the camera ready version.

Requirements

• python3.7
• pytorch>=1.5.0
• cuda10.2

Datasets

GTA5

Synthia

Cityscapes

Dark Zurich

Pretrained Models

1. The source only model for GTA5 and Synthia are provided by AdaptSegNet.
2. For day-to-night adaptation, please download the model pretrained on Cityscapes here.

Download these pretrained models and put them into the pretrained_model folder.

Training and Evaluation Instruction for GTA5->Cityscapes

Modify the all data paths in the train_config.py and test_config.py.

bash run.sh

Acknowledgment

Part of our code is from MixStyle and AdaptSegNet. We gratefully thank the authors for their great work. Also thank the authors of ASM for introducing this one-shot UDA setting.

Citation

If you think this paper is useful for your research, please cite our paper:

@inproceedings{wu2021style,
  title={Style Mixing and Patchwise Prototypical Matching for One-Shot Unsupervised Domain Adaptive Semantic Segmentation},
  author={Wu, Xinyi and Wu, Zhenyao and Lu, Yuhang and Ju, Lili and Wang, Song},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  year={2022}
}