[ECCV 2022] Meta-Sampler: Almost Universal yet Task-Oriented Sampling for Point Clouds

This is the PyTorch implementation of the paper Meta-Sampler: Almost Universal yet Task-Oriented Sampling for Point Clouds which will appear in ECCV-2022 Conference. ** The readability of the code will continue to be polished. **

Cite this work

@inproceedings{metasampler,
  title={Meta-Sampler: Almost Universal yet Task-Oriented Sampling for Point Clouds},
  author={Cheng, Ta-Ying and 
          Hu, Qingyong and 
          Xie, Qian and 
          Trigoni, Niki and 
          Markham, Andrew},
  booktitle={European Conference on Computer Vision (ECCV)},
  year={2022}
}

Preliminaries

The meta-sampler was built on top of the official PyTorch SampleNet implementation and the training algorithm is performed on pretrained point cloud networks: PointNet/PointNet2, Point Completion Network (PCN), and PCRNet. Please cite them accordingly when using their code. The essential components from SampleNet and PointNet/PCN are contained in this repository. To test on PCRNet, clone the pcrnet implementation into this github repository.

The code uses standard ModelNet40 dataset that can also be obtained here.

Training models for using joint training

To perform the proposed joint-training on a particular network task using unsampled point clouds:

python train_samplenet_networktask_ensemble.py

and replace networktask with the designated task (e.g., classification/pcn/retrieval).

Training meta-sampler

To perform meta-sampler training:

python train_samplenet_meta.py

Checkpoints can be found here.