Prototype Adaption and Projection for Few- and Zero-shot 3D Point Cloud Semantic Segmentation [pdf]

Overview

Running

Installation and data preparation please follow attMPTI.

Training

Pretrain the segmentor which includes feature extractor module on the available training set:

bash scripts/pretrain_segmentor.sh

Train our method under few-shot setting:

bash scripts/train_PAP.sh

Train our method under few-and zero-shot setting:

bash scripts/train_PAPFZ.sh

Evaluation

Test our method under zero-shot setting:

bash scripts/eval_PAPFZ.sh

Test our method under few-shot setting:

bash scripts/eval_PAP.sh

Note that the above scripts are used for 2-way 1-shot on S3DIS (S^0). Please modify the corresponding hyperparameters to conduct experiments on other settings.

Citation

Please cite our paper if it is helpful to your research:

@article{PAPFZS3D,
  title={Prototype Adaption and Projection for Few- and Zero-shot 3D Point Cloud Semantic Segmentation},
  author={He, Shuting and Jiang, Xudong and Jiang, Wei and Ding, Henghui},
  journal={IEEE Transactions on Image Processing},
  year={2023},
  publisher={IEEE}
}

Acknowledgement

We thank DGCNN (pytorch) and attMPTI for sharing their source code.