[go: up one dir, main page]
Skip to content

Latest commit

 

History

History
123 lines (89 loc) · 6.44 KB

README.md

File metadata and controls

123 lines (89 loc) · 6.44 KB
Raw

masked

masked is a cli tool made to automatically anonymize eyes in videos. It uses detectron2 for keypoints detections, and some weird math to apply a mask.

This project was inspired by one of the poster from the movie Parasite, so the attempt was to only cover the eyes (just like in the poster). If you tweak the default values, you can cover whole faces, but it might be easier for you to use a tool made for that purpose, like deface.

Parasite poster masked demo
Parasite Poster
Parasite.masked.demo.mp4

Compatibility

masked is compatible with Windows, Linux and macOS.

If you have a nvidia CUDA-enabled GPU, you can (and you really should) install CUDA to benefit from hardware acceleration, making the inference faster (see Windows or Linux CUDA installation guide).

If you don't have CUDA installed, you can still use masked, you just need to use the --cpu option.

Installation

Notes

masked uses poetry as a dependency manager, make sure it is installed before getting started (see poetry installation instructions).

It also uses ffmpeg in order to keep the audio of the videos you mask (see ffmpeg website to download a binary). FFmpeg executable needs to be in your PATH for masked to use it. You can also use the --no-audio option to skip the audio copy.

Steps

You first need to get the source of masked, and use poetry to create a virtualenv and install the dependencies:

git clone https://github.com/Aziks0/masked.git
cd masked
poetry install

You then need to install detectron2 (and pytorch + torchvision if you build it from source)

Warning

Remember that you are using poetry! Install the following package in the poetry virtualenv with poetry run <PIP_COMMAND>.

Linux

detectron2 has pre-built wheels for Linux, but you can also build it from source if you prefer.

macOS

You need to build detectron2 from source, please refer to the detectron2 installation guide.

Windows

You first need to install pytorch and torchvison. Please refer to the pytorch website for the installation guide.

Then, you need to install detectron2. It doesn't support Windows -officially-, so you need a "custom" source to build and install it. You can either use my fork version or use your own fork and fix the source yourself (see this commit).

Exemple installation with CUDA 11.3 installed:

poetry run pip install torch torchvison --extra-index-url https://download.pytorch.org/whl/cu113
poetry run pip install git+https://github.com/Aziks0/detectron2-windows.git

Please refer to the detectron2 build guide for further information.

Usage

Use poetry run masked -i <VIDEO_PATH> to run the tool. Check the available options here or with the poetry run masked -h command.

Options

usage: masked [-h] --input INPUT [--output OUTPUT] [--threshold THRESHOLD] [--mask-scale WIDTH HEIGHT]
              [--mask-color RED GREEN BLUE] [--remove-duplicates] [--faces FACES] [--no-audio] [--visualize]

Anonymize faces in videos

optional arguments:
  -h, --help            show this help message and exit
  --input INPUT, -i INPUT
                        Path to video file.
  --output OUTPUT, -o OUTPUT
                        Path to the output video file.
  --threshold THRESHOLD, -t THRESHOLD
                        Minimum score for faces to be masked, [0,1].
  --mask-scale WIDTH HEIGHT, -ms WIDTH HEIGHT
                        Scale factor for face masks, ]0,+∞].
  --mask-color RED GREEN BLUE, -mc RED GREEN BLUE
                        Face masks color, in RGB.
  --remove-duplicates, -rd
                        Try to remove duplicated detections.
  --faces FACES, -f FACES
                        Specify the number of faces that will appear in the video. Only the X most probable detections
                        will be anonymized.
  --no-audio            Don't keep the audio from the input video.
  --visualize           Draw scores and boxes. This argument is only compatible with {input|output|threshold|no-audio}
                        arguments.

Acknowledgments