ART-Point: Improving Rotation Robustness of Point Cloud Classifiers via Adversarial Rotation
Introduction
PyTorch implementation for the paper ART-Point: Improving Rotation Robustness of Point Cloud Classifiers via Adversarial Rotation (CVPR 2022).
Repository still under construction/refactoring.
Installation
Install Requirements
$ cd ART-Point/
$ conda env create -f environment.yaml
Download ModelNet40 and ShapeNet Parts
We use two datasets:
After downloading, you should convert the .txt dataset into numpy file (.npy). Then, you can use our code for training and evaluation.
You can use the codes in “https://github.com/yanx27/Pointnet_Pointnet2_pytorch/tree/master/data_utils” for pre-pocessing.
Pretraining Models
We use the folloing implemetations to respectively pretrain classifiers on ModelNet40 and ShapeNet16.
- DGCNN
- PointNet/PointNet++
After pre-training, you should move the pre-trained models into corresponding folders