Latent Diffusion Face Anonymisation LDFA

This repository contains the code for the paper LDFA: Latent Diffusion Face Anonymization for Self-driving Applications.

Structure

Dockerfile

The dockerfile is used to start container which runs the Automatic1111 web UI for stable diffusion. LDFA uses the API to conveniently use a stable diffusion model for the anonymization of human faces.

Scripts

detect_faces.py - This script uses RetinaFace to detect faces on a given dataset.
ldfa_face_anon.py - This script implements the LDFA anonymization method.
simple_face_anon.py - This script implements the naive anonymization methods cropping, gaussian noise and pixelaziation which are applied on detected faces.

Test

The tests are not meant to be used as a unit test, but to show a quick script usage of our tooling. The tests are run on some samples from the cityscapes dataset.

Usage

Please use the provided Docker container. Make sure that you have Docker Compose V2. See Diff between V1 and V2

Prior to using this tool, please make sure that you have correctly set up the image, mask, anonymized, and weights volumes inside the docker-compose.yml file. Furthermore, you can freely specify which GPU should be used. You can start the needed docker instances with docker compose up. The script will look for all images in the given root folder. The default extension is png. If you want to use other extension, you can provide a flag to the corresponding python scripts, e.g. --image_extension=jpg.

Once the docker container is running you can generate masks using:

docker compose exec anon python3 /tool/scripts/detect_faces.py --image_dir=/data/images --mask_dir=/data/masks

and anonymize the detected faces using:

docker compose exec anon python3 /tool/scripts/ldfa_face_anon.py --image_dir=/data/images --mask_dir=/data/masks --output_dir=/data/anonymized

Citation

If you are using LDFA in your research, please consider to cite us.

@InProceedings{Klemp_2023_CVPR,
    author    = {Klemp, Marvin and R\"osch, Kevin and Wagner, Royden and Quehl, Jannik and Lauer, Martin},
    title     = {LDFA: Latent Diffusion Face Anonymization for Self-Driving Applications},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
    month     = {June},
    year      = {2023},
    pages     = {3198-3204}
}