3d-isometry-robust's Introduction

[CVPR2020] On Isometry Robustness of Deep 3D Point Cloud Models under Adversarial Attacks

Environment

Ubuntu 16.04.5 LTS
GPU RTX2080ti
Python 3.7
Install the python dependencies with

pip install -r requirements.txt

Data

[ModelNet40] automatically downloaded
[ShapeNet] /fxia22/pointnet.pytorch (follow the guidence for downloading)
The default path of data is '/data'.

Usage Sample

Train model

With default parameters setting, run

python train.py --data modelnet40 --model pointnet

Trained model is stored in '/checkpoints' with log in '/logs_train'.

Launch attack

If you don't want to retrain the model, download a trained model here (with ModelNet40 data, PointNet model), move it to '/checkpoints', then run

python attack.py --data modelnet40 --model pointnet --model_path 'example'

The attack log is stored in '/logs_attack'. The attack is default to be CTRI since TSI is done at the same time.

Supplementary materials

Please check here for supplementary materials mentioned in this paper.

References

PointNet /charlesq34/pointnet, /fxia22/pointnet.pytorch
PointNet++ /charlesq34/pointnet2, /yanx27/Pointnet_Pointnet2_pytorch
DG-CNN /WangYueFt/dgcnn
RS-CNN /Yochengliu/Relation-Shape-CNN
Thompson Sampling /andrecianflone/thompson
Adversarial Attacks /MadryLab, /YyzHarry/ME-Net

3d-isometry-robust's People

Contributors

Stargazers

Watchers

3d-isometry-robust's Issues

why ModelNet40 dataset url is None ?

RuntimeError: cublas runtime error : the GPU program failed to execute at /pytorch/aten/src/THC/THCBlas.cu:441

python train.py --data modelnet40 --model dgcnn

and

python train.py --data modelnet40 --model pointnet2

Adversarial example visualization

Hi, thanks for your work, it is very interesting. I want to cite your code and i do not know how to save and visualize the adversarial point cloud, could you give me some advice? Hoping your reply.

There were some issues training the PointCNN model.

Thank you very much for sharing!

I ran the following two commands while training PointCNN:
python train.py --data shapenetpart --model pointcnn
python train.py --data modelnet40 --model pointcnn

It was found that the training speed was particularly slow, and even the Volatile GPU-Util was 0%. Is there something wrong with me?

Looking forward to your reply. :)

skywalker6174 / 3d-isometry-robust Goto Github PK