AdvMix

Official code for our CVPR 2021 paper: "When Human Pose Estimation Meets Robustness: Adversarial Algorithms and Benchmarks".

Getting started

• Installation

# clone this repo
git clone https://github.com/AIprogrammer/AdvMix
# install dependencies
pip install -r requirements
# make nms
cd AdvMix
cd lib
make
# install cocoapi
# COCOAPI=/path/to/clone/cocoapi
git clone https://github.com/cocodataset/cocoapi.git $COCOAPI
cd $COCOAPI/PythonAPI
# Install into global site-packages
make install
# Alternatively, if you do not have permissions or prefer
# not to install the COCO API into global site-packages
python3 setup.py install --user

• Download the datasets COCO, MPII, and OCHuman. Put them under "./data". The directory structure follows HRNet.

Benchmarking

Contruct benchmarking datasets

sh scripts/make_datasets.sh

Visualization examples

Benchmark results

Note: There may be small gap between the results by Evaluation and results in our paper due to randomness of operations in package 'imagecorruptions'.

AdvMix

Training

• MPII

sh scripts/train.sh mpii

• COCO

sh scripts/train.sh coco

Evaluation

sh scripts/test.sh coco
sh scripts/test.sh mpii

Quantitative results

Comparisons between standard training and AdvMix on COCO-C. For top-down approaches, results are obtained with detected bounding boxes of HRNet. We see that mPC and rPC are greatly improved, whilst clean performance AP^* can be preserved

Visualization results

Qualitative comparisons between HRNet without and with AdvMix. For each image triplet, the images from left to right are ground truth, predicted results of Standard HRNet-W32, and predicted results of HRNet-W32 with AdvMix.

Citations

If you find our work useful in your research, please consider citing:

@article{wang2021human,
  title={When Human Pose Estimation Meets Robustness: Adversarial Algorithms and Benchmarks},
  author={Wang, Jiahang and Jin, Sheng and Liu, Wentao and Liu, Weizhong and Qian, Chen and Luo, Ping},
  journal={arXiv preprint arXiv:2105.06152},
  year={2021}
}

License

Our research code is released under the MIT license. See LICENSE for details.

Acknowledgments

Thanks for open-source code HRNet.