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

May 28, 2023 ยท View on GitHub

Overview

framework

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.