PAConv: Position Adaptive Convolution with Dynamic Kernel Assembling on Point Clouds

by Mutian Xu*, Runyu Ding*, Hengshuang Zhao, and Xiaojuan Qi.

Introduction

This repository is built for the official implementation of:

PAConv: Position Adaptive Convolution with Dynamic Kernel Assembling on Point Clouds (CVPR2021) [arXiv]

If you find our work useful in your research, please consider citing:

@inproceedings{xu2021paconv,
  title={PAConv: Position Adaptive Convolution with Dynamic Kernel Assembling on Point Clouds},
  author={Xu, Mutian and Ding, Runyu and Zhao, Hengshuang and Qi, Xiaojuan},
  booktitle={CVPR},
  year={2021}
}

Highlight

All initialization models and trained models are available.
Provide fast multiprocessing training (nn.parallel.DistributedDataParallel) with official nn.SyncBatchNorm.
Incorporated with tensorboardX for better visualization of the whole training process.
Support recent versions of PyTorch.
Well designed code structures for easy reading and using.

Usage

We provide scripts for different point cloud processing tasks:

Object Classification task on Modelnet40.
Shape Part Segmentation task on ShapeNetPart.
Indoor Scene Segmentation task on S3DIS.

You can find the instructions for running these tasks in the above corresponding folders.

Performance

The following tables report the current performances on different tasks and datasets. ( * denotes the backbone architectures)

Object Classification on ModelNet40

Method	OA
PAConv (PointNet)*	93.2%
PAConv (DGCNN)*	93.9%

Shape Part Segmentation on ShapeNet Part

Method	Class mIoU	Instance mIoU
PAConv (DGCNN)*	84.6%	86.1%

Indoor Scene Segmentation on S3DIS Area-5

Method	S3DIS mIoU
PAConv (PointNet++)*	66.58%

Contact

You are welcome to send pull requests or share some ideas with us. Contact information: Mutian Xu (mino1018@outlook.com) or Runyu Ding (ryding@eee.hku.hk).

Acknowledgement

Our code base is partially borrowed from PointWeb, DGCNN and PointNet++.

MohamedAfham/PAConv