Official Implementation of PointCert

Introduction

This repo is the official implementation of PointCert in pytorch.

For any other implementation details, please refer to our paper PointCert: Point Cloud Classification with Deterministic Certified Robustness Guarantees. (CVPR 2023) [Paper]

Setup your virtual environment

• Initialize virtual environment:

  conda create -n PointCert python=3.7
  

• Install the pytorch. The latest codes are tested on Ubuntu 16.04, PyTorch 1.8.1 and Python 3.7:

  pip install torch==1.8.1
  pip install torchvision==0.9.1
  

Preparation

Download the code and ModelNet40

1. Clone this repo from the GitHub.

    git clone https://github.com/jzhang538/CVPR23-PointCert.git
   

2. Download ModelNet40 dataset [here] and save it to ./data/modelnet40_normal_resampled.

3. Finally, your directory structure should look something like this:

    CVPR23-PointCert
    └── data
        └── modelnet40_normal_resampled
        	└── modelnet40_shape_names.txt
        	└── modelnet40_provider.txt
        	└── ...
        	└── class1 (e.g., xbox)
        	└── class2 (e.g., wardrobe)
        	└── ...
   

Preprocess ModelNet40

Use the following command to preprocess ModelNet40 dataset. Specifically, we divide each point cloud into sub-point clouds based on the MD5 hash values. By default, the number of sub-point clouds for each point cloud is 400. (The argument num-group indicates m in our paper)

python3 preprocess.py --num_group 400

Compile distance libraries for PCN:

cd distance
python3 setup.py install

Run the PointCert (Scenario II)

1. Train a base point cloud classifier provided by a model provider using sub-point clouds:

    python3 scenarioII/train_provider_scenarioII.py --num_group 400
   

2. Certify the base point cloud classifier in Scenario II:

    python3 scenarioII/test_scenarioII.py --num_group 400 --use_ensemble \
        --classifier_path ./scenarioII/provider/best_model_m_400.pth
   

Run the PointCert (Scenario III)

1. Train a base point cloud classifier provided by a model provider in a standard manner (using point cloud or subsampled point cloud, e.g., 1024 points):

    python3 scenarioIII/train_provider_scenarioIII.py --num_point 1024
   

2. Train a PCN using unlabeled (or partially labeled) customer point clouds (e.g. PCN takes sub-point clouds with 32 points as input to output a reconstructed point cloud with 1024 points):

    python3 scenarioIII/train_customer_scenarioIII.py --num_group 400 \
        --num_input 32 --num_recon 1024 --ratio 0.25 --l 0.0005 \
        --classifier_path ./scenarioIII/provider/best_model_standard_1024.pth
   

3. Certify the base point cloud classifier in Scenario III:

    python3 scenarioIII/test_scenarioIII.py --num_group 400 --use_ensemble \
        --classifier_path ./scenarioIII/provider/best_model_standard_1024.pth \
        --pcn_path ./scenarioIII/customer/best_model_m_400_input_32_recon_1024.pth
   

Draw certified accuracy curves

Use the following command and the results are saved at './curves' folder.

python3 draw_certified_accuracy_curve.py --scenario II
python3 draw_certified_accuracy_curve.py --scenario III

Pretrained Models

Download pre-trained models [here] and put them to corresponding folders.

Note:

• Please skip sub-point clouds with extremely small number of points (e.g., <4) during training when num-group in scenarioII is large!!!

Contact us

If you have any issues with the code, please contact to this email: jinghuai.zhang@duke.edu

Citation

If you find our work useful for your research, please consider citing the paper

@article{zhang2023pointcert,
  title={PointCert: Point Cloud Classification with Deterministic Certified Robustness Guarantees},
  author={Zhang, Jinghuai and Jia, Jinyuan and Liu, Hongbin and Gong, Neil Zhenqiang},
  journal={arXiv preprint arXiv:2303.01959},
  year={2023}
}