Dense-Resolution Network for Point Cloud Classification and Segmentation

This repository is for Dense-Resolution Networ (DRNet) introduced in the following paper

Shi Qiu Saeed Anwar, Nick Barnes
"Dense-Resolution Network for Point Cloud Classification and Segmentation"
IEEE/CVF Winter Conference on Applications of Computer Vision (WACV 2021)

Paper

The paper can be downloaded from here (arXiv) or here (CVF), together with the supplementary material.

Motivation

Implementation

Python 3.6
Pytorch 1.3.0
Cuda 10.0

Dataset

Download the ShapeNet Part Dataset and upzip it to your rootpath. Alternatively, you can modify the path of your dataset in cfgs/config_partseg_gpus.yaml and cfgs/config_partseg_test.yaml.

CUDA Kernel Building

For PyTorch version <= 0.4.0, please refer to Relation-Shape-CNN.
For PyTorch version >= 1.0.0, please refer to Pointnet2_PyTorch.

Note:
In our DRNet, we use Farthest Point Sampling (e.g., pointnet2_utils.furthest_point_sample) to down-sample the point cloud. Also, we adpot Feature Propagation (e.g., pointnet2_utils.three_nn and pointnet2_utils.three_interpolate) to up-sample the feature maps.

Training

sh train_partseg_gpus.sh

Due to the complexity of DRNet, we support Multi-GPU via nn.DataParallel. You can also adjust other parameters such as batch size or the number of input points in cfgs/config_partseg_gpus.yaml, in order to fit the memory limit of your device.

Voting Evaluation

You can set the path of your pre-trained model in cfgs/config_partseg_test.yaml, then run:

sh voting_test.sh

Citation

If you find our paper is useful, please cite:

@inproceedings{qiu2021dense,
  title={Dense-Resolution Network for Point Cloud Classification and Segmentation},
  author={Qiu, Shi and Anwar, Saeed and Barnes, Nick},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  month={January},
  year={2021},
  pages={3813-3822}
}

Acknowledgement

The code is built on Pointnet2_PyTorch, Relation-Shape-CNN, DGCNN. We thank the authors for sharing their codes.

ShiQiu0419 / DRNet

readme