Installation and data preparation please follow attMPTI.
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
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.
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}
}
We thank DGCNN (pytorch) and attMPTI for sharing their source code.