BLANKET: Anonymizing Faces in Infant Video Recordings

ICDL 2025

click here for full video

       

📋 Overview

BLANKET is a two-stage pipeline for seamless, expression-preserving face anonymization in infant videos. It replaces identities with synthetic baby faces while maintaining gaze, head pose, and emotional expression, enabling ethical data sharing and robust downstream analytics.

Key contributions:

• Two-stage design: diffusion-based inpainting + temporally-consistent swap
• Attribute preservation: expression, gaze, head orientation, eye/mouth openness
• High downstream performance: ~90% detection AP, ~97% pose estimation retention
• Outperforms SOTA: beats DeepPrivacyV2 on de-identification, perceptual metrics and downstream task performance

Warning

This code is experimental and not yet production-ready. Anonymization results may be imperfect and may miss detections. Full reliability will be achieved once the “missing detections” problem is solved (see Roadmap below).

📢 News

• May 2025: Paper accepted to ICDL 2025! 🎉

🚀 Installation

To install BLANKET and its dependencies:

git clone https://github.com/ctu-vras/blanket-infant-face-anonym.git
cd blanket-infant-face-anonym
pip install -r requirements.txt
pip install .

Pre-trained models

• YOLOv11L-face - download from Ultralytics

🎮 Demo

To run the image anonymization demo:

1. Set your input and output folders in blanket/configs/config.yaml:

• input_folder: path to your images (default: ../data/)
• output_folder: path for results (default: outputs/)

2. Run the demo script:

python run_image_anonymization.py

This will process all images in your input folder, apply face detection and three anonymization methods (black box, pixelation, gaussian blur), and save composite results to your output folder.

You can adjust detection and anonymization settings in blanket/configs/config.yaml.

📊 Evaluation

BLANKET outperforms DeepPrivacy V2 in all measured metrics.

Metric	BLANKET	DeepPrivacy2
Identity cosine distance (↓)	0.11 ± 0.18	0.19 ± 0.26
Emotion preservation (↑)	0.51 ± 0.13	0.27 ± 0.11
Temporal landmark corr. (↑)	0.956 ± 0.064	0.860 ± 0.140
Detection AP vs. orig. (↑)	90.7 mAP	81.5 mAP
Pose AP vs. orig. (↑)	97.2 mAP	79.1 mAP

🗺️ Roadmap

We will continue refining BLANKET with a focus on quality and reliability:

• Implement Stable Diffusion–based inpainting
• Implement FaceFusion in video and a video demo
• Ensure robust anonymization in frames where faces are not detected

🙏 Acknowledgments

Supported by GA CR 25-18113S, EC Digital Europe CEDMO 2.0 101158609, CTU SGS23/173/OHK3/3T/13. Thanks to Adéla Šubrtová for early feedback. Thanks to Max Familly Fun for the banner picture.

📝 Citation

If you use BLANKET, please cite:

@inproceedings{hadera2025BLANKET,
  author = {Hadera, Ditmar and Cech, Jan and Purkrabek, Miroslav and Hoffmann, Matej},
  booktitle = {2025 IEEE International Conference on Development and Learning (ICDL)},
  month = sep,
  pages = {1--8},
  title = {{BLANKET: Anonymizing Faces in Infant Video Recordings}},
  year = {2025}
}