Check out the new DeepPrivacy2! It significantly improves anonymization quality compared to this repository.
DeepPrivacy is a fully automatic anonymization technique for images.
This repository contains the source code for the paper "DeepPrivacy: A Generative Adversarial Network for Face Anonymization" published at ISVC 2019, and "Image Inpainting with Learnable Feature Imputation " published at GCPR 2020.
The DeepPrivacy GAN never sees any privacy sensitive information, ensuring a fully anonymized image.
It utilizes bounding box annotation to identify the privacy-sensitive area, and sparse pose information to guide the network in difficult scenarios.
DeepPrivacy detects faces with state-of-the-art detection methods.
Mask R-CNN is used to generate a sparse pose information of the face, and DSFD is used to detect faces in the image.
Check out the original version here:
Check out GCPR readme to reproduce our results from our publication "Image Inpainting with Learnable Feature Imputation".
- FP16 inference
- Support for face detection and landmark detection with a single SSD model (retinanet).
- Face alignment before anonymization - improves performance for rotate faces.
- Heavily refactored code.
- All improvements from our paper "Image Inpainting with Learnable Feature Imputation" are included.
- Support for standard image inpainting datasets (CelebA-HQ and Places2).
- Support for video inference
Install the following:
- Pytorch >= 1.7.0 (Some checkpoints requires a bug fix from pytorch 1.7 - the current master branch)
- Torchvision >= 0.6.0
- NVIDIA Apex (If you want to train any models - nolt needed for inference)
- Python >= 3.6
Simply by running our setup.py file:
python3 setup.py install
or with pip:
pip install git+https://github.com/hukkelas/DeepPrivacy/
In our experiments, we use docker as the virtual environment.
Our docker image can be built by running:
cd docker/
docker build -t deep_privacy . We have a file command line interface to anonymize images.
python3 anonymize.py -s input_image.png -t output_path.png
You can change the model with the "-m" or "--model" flag (see model zoo). The cli accepts image files, video files, and directories.
The cli is also available outside the folder python -m deep_privacy.cli.
Also check out python -m deep_privacy.cli -h for more arguments.
Test out the model with
python webcam.py
Also check out python webcam.py -h for more arguments.
| Model | Dataset | Detector | Num parameters |
|---|---|---|---|
| deep_privacy_v1 | FDF 128x128 | SSD-based RetinaNet for face detection + Mask RCNN for keypoints | 46.92M |
| fdf128_rcnn512 (recommended) | FDF 128x128 | SSD-based RetinaNet for face detection + Mask RCNN for keypoints | 47.39M |
| fdf128_retinanet512 | FDF 128x128 | SSD-based RetinaNet with ResNet50 backbone | 49.84M |
| fdf128_retinanet256 | FDF 128x128 | SSD-based RetinaNet with ResNet50 backbone | 12.704M |
| fdf128_retinanet128 | FDF 128x128 | SSD-based RetinaNet with ResNet50 backbone | 3.17M |
Combined keypoint and face detection in one single-shot model (SSD-based). Uses five landmarks: eyes, nose, and mouth (left and right). Model is based on RetinaFace. The model is significantly faster, but has poorer detection than DSFD.
Uses Mask R-CNN is for keypoint detection, and DSFD for face detection.
If you find this code useful, please cite the following:
@InProceedings{10.1007/978-3-030-33720-9_44,
author="Hukkel{\aa}s, H{\aa}kon
and Mester, Rudolf
and Lindseth, Frank",
title="DeepPrivacy: A Generative Adversarial Network for Face Anonymization",
booktitle="Advances in Visual Computing",
year="2019",
publisher="Springer International Publishing",
pages="565--578",
isbn="978-3-030-33720-9"
}
The FDF dataset will be released at github:hukkelas/FDF
All code is under MIT license, except the following.
Code under deep_privacy/detection:
- DSFD is taken from https://github.com/hukkelas/DSFD-Pytorch-Inference and follows APACHE-2.0 License
- Mask R-CNN implementation is taken from Pytorch source code at pytorch.org
- FID calculation code is taken from the official tensorflow implementation: https://github.com/bioinf-jku/TTUR
- Checkpoints published with the model follows the license of the FDF dataset.