README.md

NP-SemiSeg

NP-SemiSeg: When Neural Processes meet Semi-Supervised Semantic Segmentation

Jianfeng Wang¹, Daniela Massiceti², Xiaolin Hu³, Vladimir Pavlovic⁴ and Thomas Lukasiewicz¹

University of Oxford¹, Microsoft Research², Tsinghua University³, Rutgers University⁴

In ICML 2023

Build

please run with the following command:

conda env create -f NP-SemiSeg.yaml
conda activate NP-SemiSeg

Experiment

We release the neural processes header (np_head.py) for semi-supervised semantic segmentation, and how it is used is shown in the two segmentation frameworks, namely U2PL and AugSeg.

Please download pretrained resnet50.pth and datasets at first. The datasets can be found in the original U2PL and AugSeg repos. Here we provide download links of resnet50.pth for convenience.

ResNet-50:

Google Drive	Baidu Disk
download	download (code: mhbx)

Please put the resnet50.pth in the "pretrained" directory and datasets in the "data" directory, in both U2PL and AugSeg.

For U2PL:

cd experiments/cityscapes/744/np/

sh train.sh <num_gpu> <port>

After training, the model should be evaluated by

sh eval.sh

For AugSeg:

Please configure your yaml file in a running script "./scripts/run_abls_citys.sh", and then run:

sh ./scripts/run_abls_citys.sh

Citation

@inproceedings{wang2023np,
  title={NP-SemiSeg: when neural processes meet semi-supervised semantic segmentation},
  author={Wang, Jianfeng and Massiceti, Daniela and Hu, Xiaolin and Pavlovic, Vladimir and Lukasiewicz, Thomas},
  booktitle={International Conference on Machine Learning},
  pages={36138--36156},
  year={2023},
  organization={PMLR}
}

Acknowledgement

The released codes contain the contents from U2PL and AugSeg. We thank for their contributions.