Learning to Censor by Noisy Sampling

This is the code for ECCV Submission - Paper ID: 3755.

Installation

We strongly recommend working with Docker Engine and Nvidia-Docker. At this moment, the container can only run on a CUDA (linux/amd64) enabled machine due to specific compiled ops from Pointnet2_PyTorch.

Pull and run the Docker container

docker pull asafmanor/pytorch:samplenetreg_torch1.4
docker run --runtime nvidia -v $(pwd):/workspace/ -it --name samplenetreg asafmanor/pytorch:samplenetreg_torch1.4

Alternatively, build your own Docker image

On the host machine

docker build -t samplenetreg_torch1.4_image .
docker run --runtime nvidia -v $(pwd):/workspace/ -it --name samplenetreg samplenetreg_torch1.4_image

Inside the Docker container

cd /root/Pointnet2_PyTorch
git checkout 5ff4382f56a8cbed2b5edd3572f97436271aba89
pip install -r requirements.txt
pip install -e .
cd /workspace

Usage

Data preparation

FaceScape

Download the facescape data from this website to a desired folder to read from. Run data/Preprocess_Data_v2.ipynb to sample pointclouds (.npy) from mesh (.obj) files and generate annotations for the same. Kindly change the read/write paths in the notebook for proper execution.

Training and evaluating

Facescape

For a quick start please use the scripts provided in scripts/ directory. Run all scripts from the root directory of this repository. Note that a temporary --base-name is added to all scripts to maintain anonymity and needs be modified before running the scripts (parent directory containing annotation file and npy data after preprocessing).

1. Train base pointnet classifiers + finetune samplers: bash ./facescape/scripts/baseline/gender.sh
bash ./facescape/scripts/baseline/exp.sh
   

2. Train with proxy attacker model for privacy + finetune worst case attacker: a. CBSN (Our Method) bash ./facescape/scripts/cbns/exp_gender_best.sh
   b. Line Cloud bash ./facescape/scripts/line/exp_gender_best.sh
   c. AS-AN bash ./facescape/scripts/as_an/exp_gender_best.sh
   d. AS-ON bash ./facescape/scripts/as_on/exp_gender_best.sh
   e. OS-AN bash ./facescape/scripts/os_an/exp_gender_best.sh
   

Results for the experiment will be collected in the provided log-directory (eg: ./facescape/log/cbns/finetune/test.log). In order to reproduce the normalized hypervolume, a sweep over the hyperparameters to regulate privacy-utility is required which may take a long time to compute. For those interested, the parameters can easily be regulated by modifying the arguments the bash scripts (./facescape/scripts/).

The resulting point clouds can be visualized using results/plot_pointclouds.ipynb with appropriate paths to trained model files.

If you use our work in your project kindly cite:

@article{chopra2022learning,
  title={Learning to Censor by Noisy Sampling},
  author={Chopra, Ayush and Java, Abhinav and Singh, Abhishek and Sharma, Vivek and Raskar, Ramesh},
  journal={arXiv preprint arXiv:2203.12192},
  year={2022}
}

Acknowledgment

This code builds upon the code provided in Samplenet, PointNetLK, Pointnet2_PyTorch and KNN_CUDA. We thank the authors for sharing their code.