README.md

Anonymization for Skeleton Action Recognition

This repository is the official implementation of 'Anonymization for Skeleton Action Recognition' (AAAI2023)

Anonymization framework

Prerequisites

• Python3
• Pytorch
• Run pip install -r requirements.txt for installing other python libraries
• We use Wandb for experiment tracking

Compile cuda extensions

cd ./model/Temporal_shift
bash run.sh

Data Preparation

• We use NTU RGB+D skeleton-only datasets (nturgbd_skeletons_s001_to_s017.zip).
• After downloading datasets, generate the skeleton data with this command.

python data_gen/ntu_gendata.py --data_path <path to nturgbd+d_skeletons>

Training

To train the models in the paper, run this command:

python main.py --config ./config/train_adver_resnet.yaml
python main.py --config ./config/train_adver_unet.yaml

Pre-trained Models

We provide two pre-trained model with NTU60. You can download pretrained models here:

• Google Drive

Model	Anonymizer network	Re-iden. acc.	Action acc.
./save_models/pretrained_resnet.pt	ResNet	4.20%	91.75%
./save_models/pretrained_unet.pt	UNet	5.70%	91.45%

To test the pre-trained models given above, run this command:

python main.py --config ./config/train_adver_resnet.yaml
python main.py --config ./config/train_adver_unet.yaml

Also, we provide more privacy pre-trained models for test privacy model. You can find at /save_models/ntu_pretrained_x.

Acknowledgements

This code is based on Shift-GCN. Also, we use U-Net for anonymizer network. Thanks to the original authors!☺️