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).

Installation

Install Requirements

$ cd ART-Point/
$ conda env create -f environment.yaml

Download ModelNet40 and ShapeNet Parts

We use two datasets:

• ModeNet40
• ShapeNet16

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 at "./pretrained_models/"

Train and Evaluate

ModelNet40

To train and evaluate ART-Point with one-step optimization on ModelNet40 using PointNet backends run:

$ python train_classification_onestep.py --angles 1 --batch_size 17 --inner_epoch 200 --iters 10 --log_dir pn1_onestep --rp

To train and evaluate ART-Point with iterative optimization on ModelNet40 using PointNet backend run:

$ python train_classification_dynamic.py --angles 1 --batch_size 17 --epoch 50 --inner_epoch 50 --iters 10 --log_dir pn1_dynamic --rp

ShapeNet16

To train and evaluate ART-Point with one-step optimization on ShapeNet16 using PointNet backends run:

$ python train_classification_onestep_s16.py --angles 1 --batch_size 17 --inner_epoch 200 --iters 10 --log_dir pn1_onestep_s16 --rp

Contact

You are welcome to send pull requests or share some ideas with us. Contact information: Robin Wang (robin_wang@pku.edu.cn).