We propose a new neural network module dubbed EdgeConv suitable for CNN-based high-level tasks on point clouds including classification and segmentation. EdgeConv is differentiable and can be plugged into existing architectures.
DGCNN
is the author's re-implementation of Dynamic Graph CNN, which achieves state-of-the-art performance on point-cloud-related high-level tasks including category classification, semantic segmentation and part segmentation.
Further information please contact Yue Wang and Yongbin Sun.
The classification experiments in our paper are done with the pytorch implementation.
- pytorch-geometric
- pytorch-dgcnn (This implementation on S3DIS achieves significant better results than our tensorflow implementation)
The performance is evaluated on ModelNet-C with mCE (lower is better) and clean OA (higher is better).
Method | Reference | Standalone | mCE | Clean OA |
---|---|---|---|---|
PointNet | Qi et al. | Yes | 1.422 | 0.907 |
DGCNN | Wang et al. | Yes | 1.000 | 0.926 |
- DGCNN has been successfully applied to ParticalNet in Large Hadron Collider (LHC).
Please cite this paper if you want to use it in your work,
@article{dgcnn,
title={Dynamic Graph CNN for Learning on Point Clouds},
author={Wang, Yue and Sun, Yongbin and Liu, Ziwei and Sarma, Sanjay E. and Bronstein, Michael M. and Solomon, Justin M.},
journal={ACM Transactions on Graphics (TOG)},
year={2019}
}
MIT License
The structure of this codebase is borrowed from PointNet.