Face Blurring from Video

This project provides a set of Python scripts to extract frames from a video, detect faces in each frame using RetinaFace, apply a blur effect to the detected faces, and then reconstruct a new video from the blurred frames.

---

📦 Setup and Dependencies

🧰 Requirements

• ffmpeg – Used for video frame extraction and rebuilding.
• RetinaFace (via DeepFace) – For accurate face detection.

---

🧪 ffmpeg Installation

Make sure ffmpeg is installed on your system:

• Ubuntu:

  sudo apt-get install ffmpeg

• macOS:

  brew install ffmpeg

🐍 Python Dependencies

Install required packages:

pip install -r requirements.txt

• Note: On first run, deepface will automatically download the retinaface.h5 model and store it at: ~/.deepface/weights/retinaface.h5

⚡️ GPU Acceleration (Optional)

If you’re using PyTorch and have a CUDA-enabled GPU, you can enable GPU acceleration for face detection.

For PyTorch users: Add this to the top of face_coords.py:

import torch
torch.set_default_tensor_type('torch.cuda.FloatTensor')

For retinaface-torch users: Pass the parameter device='cuda' when initializing the detector.

For retina-face users(as we are): Verify TensorFlow GPU detection: After installation, open a Python interpreter and run:

import tensorflow as tf
print(tf.config.list_physical_devices('GPU'))

Explore the tensorflow-directml-plugin if you want to use a non-CUDA GPU or prefer not to use WSL2. TensorFlow-DirectML-Plugin: This allows TensorFlow to leverage DirectX 12-compatible GPUs (including AMD, Intel, and NVIDIA GPUs) on Windows. It's an alternative to CUDA for Windows users who don't want to use WSL2. You would typically install tensorflow-cpu and then the tensorflow-directml-plugin.

🗂 Project Structure

• ├── blur_faces.py
  Main script to run the full pipeline
• ├── face_coords.py
  Face detection using RetinaFace
• ├── video_to_frames.py
  Frame extraction from video
• ├── requirements.txt Python dependencies
• └── README.md This file

🔍 Function Details

blur_faces.py

map_face(input_video, output_directory, bboxes_file, blured_frame_dir)

• Extracts frames, detects faces, saves coordinates to a JSON file, and applies blur.

apply_blur_to_image(frame_path, BLURRED_FRAMES_DIR, file, coordinates)

• Applies Gaussian blur to face regions in an image.

blur_from_bboxes(bboxes_file, output_dir, BLURRED_FRAMES_DIR)

• Reads bounding boxes from JSON and blurs corresponding faces.

open_bboxes_file_per_frame(bboxes_file)

• Generator that yields face coordinates per frame.

face_coords.py

get_face_coordinates(image_path)

• Detects faces in an image and returns a list of dictionaries with bounding box coordinates: x1, y1, x2, y2.

video_to_frames.py

extract_frames(input_video, output_directory)

• Uses ffmpeg to extract frames from a video and saves them as JPEGs.

rebuild_video_from_blurred_frames(BLURRED_FRAMES_DIR, output_video, FPS=30)

• Rebuilds video from blurred image frames using ffmpeg.

remove_directory(directory)

• Deletes a directory and all its contents.

vid_ffmpeg_to_frames_and_back.py

to use ffmpeg as library (more pythonic way) must install Python library

   pip install ffmpeg-python

consider using the functions below

• extract_frames_with_ffmpeg_library(input_video, output_directory)
• rebuild_video_from_blurred_frames_with_ffmpeg_library(BLURRED_FRAMES_DIR, output_video, FPS=30)

🧠 Credits

• DeepFace

• RetinaFace

• FFmpeg

📜 License

MIT License