desenat's Introduction

Improving Corruption Robustness: Desensitizing the Neural Network to 3D Point Cloud through Adversarial Training

State

This work is under submission to "Pattern Recognition".

Prerequisites

Install necessary packages using:

bash install.sh

Install PyGeM

cd PyGen
python setup.py install
cd ..

Data

Download ModelNet40, ModelNet40-C, and PointCloud-C datasets and put them in the Data directory:

DenAT/
└── data/
    ├── ModelNet40Ply2048
    ├── modelnet40_c
    └── pointcloud_c

Generate or download the shapely value corresponding to modelNet

Generate the shapely value corresponding to modelNet：

bash ./script/gen_shapely.py

Or you can also download it directly from this link. You need to put the file in the specified directory：

DenAT/
└── data/
    └── shaply_ModelNet40Ply2048
        └── PointNet2Encoder_ST
            ├──train
            └──test

Script

Train PointNet++ model using ST method:

bash ./script/trainST.py

Train PointNet++ model using DesenAT method:

bash ./script/train.py

Testing in ModelNet40-C

bash ./script/test_modelnetC.py

Testing in PointCloud-C

bash ./script/test_pointcloudC.py

Acknowledgment

This repository is built on reusing codes of OpenPoints and PointNeXt. We integrated APES and PointMetaBase into the code. We also have integrated methods for handling corrupted point clouds into our code, thanks to the excellent work of ModelNet-C and PointCloud-C.

Recommend Projects