jo-wang / ldbe

"Domain Adaptive Semantic Segmentation without Source Data" (ACM MM 2021)

LDBE

Pytorch implementation for two papers:

"Domain Adaptive Semantic Segmentation without Source Data", ACM MM2021. Paper link: http://arxiv.org/abs/2110.06484

"Challenging Source-free Domain Adaptive Semantic Segmentation", submitted to TPAMI.

Method

Result

GTA5 -> Cityscapes:

Methods	Source-only	LD	LDBE
mIoU	35.7	45.5	49.2

SYNTHIA -> Cityscapes:

Methods	Source-only	LD	LDBE
mIoU (16-classes)	32.5	42.6	43.5
mIoU (13-classes)	37.6	50.1	51.1

Data

Download GTA5.

Download SYNTHIA. Please use SYNTHIA-RAND-CITYSCAPES

Download Cityscapes.

Make sure the data path is consistent with the path in config file.

Training

Stage 0: Training on the source domain data.

Run "run_so.py". The trained model is available at google drive.

Stage 1: Label denoising (both positive learning and negative learning).

Set method:"ld" in config/ldbe_config.yml. Then, run "run.py". The trained model is available at google drive.

Stage 2: Boundary enhancement

Set method:"be" in config/ldbe_config.yml. Then, run "run.py". The trained model is available at google drive.

Citation

If you find our repository is helpful, please consider citing our paper

  @inproceedings{you2021domain,
  title={Domain Adaptive Semantic Segmentation without Source Data},
  author={You, Fuming and Li, Jingjing and Zhu, Lei and Chen, Zhi and Huang, Zi},
  booktitle={Proceedings of the 29th ACM International Conference on Multimedia},
  pages={3293--3302},
  year={2021}

}

Acknowledgement

https://github.com/Solacex/CCM

https://github.com/yzou2/CRST

Contact

[email protected]