DeepPrivacy2 - A Toolbox for Realistic Image Anonymization

[PDF] [Video Demo]

DeepPrivacy2 is a toolbox for realistic anonymization of humans, including a face and a full-body anonymizer.

DeepPrivacy2 detects and anonymizes individuals via three detection and synthesis networks; (1) a CSE-guided generator for individuals detected with dense pose (by CSE), (2) an unconditional full-body generator for cases where CSE fails to detect (note the segmented persons without color-coded CSE detections), and (3) a face generator for the remaining individuals (marked in red).

What's new

This repository improves over the original DeepPrivacy repository with the following new features:

• Full-body anonymization: Anonymize the entire human body with a single generator
• Improved Face Anonymization: Improved quality and higher resolution (256x256 vs. 128x128) face anonymization without relying on facial landmark detection.
• Attribute Guided Anonymiation: Anonymize faces guided on text prompts using StyleMC.
• Code cleanup and general improvements: Extensive refactoring, bugfixes, and improvements yielding improved results and faster training.

Installation

Requirements

• Pytorch >= 1.10
• Torchvision >= 0.12
• Python >= 3.8
• CUDA capable device for training. Training was done with 1-8 32GB V100 GPUs.

Installation We recommend to setup and install pytorch with anaconda following the pytorch installation instructions.

1. Clone repository: git clone https://github.com/hukkelas/deep_privacy2/.
2. Install using setup.py:

pip install -e .

or:

pip install git+https://github.com/hukkelas/deep_privacy2/

Anonymization

anonymize.py is the main script for anonymization.

The typical usage is

python3 anonymize.py configs/anonymizers/FB_cse.py -i path_to_image.png

where the first argument is the chosen anonymizer (see below for the different models) and the second a path to an image/folder/video.

There are several optional arguments, see python3 anonymize.py --help for more info.

python3 anonymize.py -h
Usage: anonymize.py [OPTIONS] CONFIG_PATH

  config_path: Specify the path to the anonymization model to use.

Options:
  -i, --input_path PATH           Input path. Accepted inputs: images, videos,
                                  directories.
  -o, --output_path PATH          Output path to save. Can be directory or
                                  file.
  --visualize                     Visualize the result
  --max_res INTEGER               Maximum resolution  of height/wideo
  --start-time, --st INTEGER      Start time (second) for vide anonymization
  --end-time, --et INTEGER        End time (second) for vide anonymization
  --fps INTEGER                   FPS for anonymization
  --detection-score-threshold FLOAT RANGE
                                  Detection threshold, threshold applied for
                                  all detection models.  [0<=x<=1]
  --visualize-detection           Visualize only detections without running
                                  anonymization.
  --multi-modal-truncation, --mt  Enable multi-modal truncation proposed by:
                                  https://arxiv.org/pdf/2202.12211.pdf
  --no-cache                      Disable loading of detection cache. Will
                                  rerun all detections.
  --amp                           Use automatic mixed precision for generator
                                  forward pass
  -t, --truncation_value FLOAT RANGE
                                  Latent interpolation truncation value.
                                  [0<=x<=1]
  --track                         Track detections over frames. Will use the
                                  same latent variable (z) for tracked
                                  identities.
  --seed INTEGER                  Set random seed for generating images.
  --person-generator PATH         Config path to unconditional person
                                  generator
  --cse-person-generator PATH     Config path to CSE-guided person generator
  --webcam                        Read image from webcam feed.
  --help                          Show this message and exit.

Singe image anonymization

python3 anonymize.py configs/anonymizers/FB_cse.py -i path_to_image.png --output-path output.png

Folder anonymization

If a folder is given as the input, all image and video files in the given folder will be anonymized and placed under --output_path. The script will duplicate the directory structure/filenames in the given folder for the output.

python3 anonymize.py configs/anonymizers/FB_cse.py path/to/input/folder --save --output_path output_folder

Video anonymization

python3 anonymize.py configs/anonymizers/FB_cse.py -i path_to_video.mp4 --output_path output.mp4

Webcam anonymization

python3 anonymize.py configs/anonymizers/FB_cse.py --webcam

Available anonymization models

DeepPrivacy2 provides the following anonymization models:

• configs/anonymizers/FB_cse.py: Full-body anonymizer that only anonymizes individuals detected by CSE. This provides the highest quality anonymization, however, some individuals might not be detected by CSE.
• configs/anonymizers/FB_cse_mask.py: Full-body anonymizer that anonymizes all individuals detected by CSE or Mask R-CNN. In difference from configs/anonymizers/FB_cse.py, this model anonymizes individuals not detected by CSE with an unguided generator.
• configs/anonymizers/FB_cse_mask_face.py: Full-body and face anonymizer that anonymizes all individuals detected by CSE, Mask R-CNN or by face detection. Compared to configs/anonymizers/FB_cse_mask.py, this model anonymizes individuals not detected by CSE or Mask R-CNN with a face anonymizer.
• configs/anonymizers/face.py: The face anonymizer only anonymizes a center crop of the face.
• configs/anonymizers/face_fdf128.py: Same as configs/anonymizers/face.py, but the generator is trained on lower resolution images (128x128 or lower). Recommended to use if you will not anonymize any faces larger than 128x128. Model will be released soon.

Attribute guided anonymization

DeepPrivacy2 allows for controllable anonymization through text prompts by adapting StyleMC. StyleMC finds global semantically meaningful directions in the GAN latent space by manipulating images towards a given text prompt with a CLIP-based loss.

The repository includes a demo for interactive text-guided anonymization. To use the demo, first:

1. Download the FDF256 dataset (see below). Only the validation set is required.
2. Run the following:

python3 attribute_guided_demo.py

The script will spin up a local webserver.

Training

First, download dataset for training (see below).

To start training, type the following:

python3 train.py configs/fdh/styleganL.py

The training automatically logs to wandb.

Model development utility scripts

Dataset inspection: To inspect the training dataset, you can use:

python3 -m tools.inspect_dataset configs/fdh/styleganL.py

Sanity check:

python3 -m tools.dryrun configs/fdh/styleganL.py

Output visualization: To visualize output of trained models:

python3 -m tools.show_examples configs/fdh/styleganL.py

Calculating metrics

python3 validate.py configs/fdh/styleganL.py

NOTE: The metrics calculated with validate.py will slightly differ from training metrics, as validate.py disables automatic mixed precision.

Dataset Setup

Setting Data directory: The default dataset directory is ./data. If you want to change the dataset directory, set the environment variable BASE_DATASET_DIR. For example, export BASE_DATASET_DIR=/work/data/.

FDF256

Follow the instructions here to download the FDF256 dataset. The dataset should be placed in the directory: data/fdf256.

FDH

Follow the instructions here to download the FDH dataset. The dataset should be placed in the directory: data/fdh.

License

This repsitory is released under Apache 2.0 License, except for the following:.

• Code under sg3_torch_utils/. This code is modified from github.com/NVlabs/stylegan2-ada-pytorch. Separate license is attached in the directory.
• Detection network: See Detectron2 License.
• All checkpoints follow the license of the datasets. See the respective datasets for more information.