Privacynator

Privacynator is an utility for removing Personally Identifiable Information (PII) from videos. It is primarily meant for videos captured from cars.

Privacynator uses the Detectron2 model from facebook research: https://github.com/facebookresearch/detectron2.

Installation

The installation process is not completely straight forward as it depends on the CUDA version of your system.

Create a virtual environment (optional, but recommended)

To create a virtual environment with python, use the following command:

python3 -m venv venv

Then, activate this environment by doing

source ./venv/bin/activate

Installing the correct dependencies

First, install Torch and Torchvision by following the instructions at https://pytorch.org/, making sure that your version of CUDA is supported.

Then, install Detectron2 by following instructions at https://detectron2.readthedocs.io/en/latest/tutorials/install.html. NOTE: The Detectron2 version must match the PyTorch version!

Finally, install the rest of the dependencies by doing

pip install -r requirements.txt

Usage

To run Privacynator, use the follwoing command:

python3 privatize.py <video-operation> <input-file-path> <output-file-path>

Where video-operation can be either reject, blur, mask or demo. More information on those can be found further down in the readme.

privatize.py also comes with several keyword arguments. For example, if you want to:

• Overwrite (-o) the output file if it exists
• Start (-s) from frame 68
• Limit (-l) video to the 419 next frames

Your command would be

python3 privatize.py -o -s 68 -l 419 <video-operation> <input-file-path> <output-file-path>

It is also possible to use different Coco classes (-c) for the application. So if you for some reason would like to blur all frisbees and giraffes in your video, you can run the command:

python3 privatize.py blur <input-file-path> <output-file-path> -c 33 24

A list of all the Coco classes can be found by following this link. Note: The Coco class numbers in the link are one-indexed, while the input Coco class numbers to privatize.py has to be 0-indexed.

Video operations

There are several possible techniques to remove PII from a video, and Privacynator supports some of them. The available methods are each implemented as its own "video operation".

The currently available video operations are:

Reject

Rejects frames from the input video containing PII. This means your output video probably will be a lot shorter than the input video.

Mask

Create black masks over regions containing PII in the video.

Blur

Blur areas containing PII in the video.

Demo

This video operation just runs Detectron2's demo visualization on every frame of the video. This can be nice for debugging if the other video operations makes weird results.

Adding a new video operation

A video operation is fairly simple. It takes an image as input and ouputs 0, 1 or more images to an output based on the content of the input image.

To add a new video operation, create a subclass of the BaseVideoOperation class and implement the function do_operation. Then, register the new video operation in the video_operation_options dict in the file video_operation_options.py.

The BaseVideoOperation class can be found at video_operations/base_video_operation.py.

Testing out Privacynator

An example of a video to privatize can be found at https://www.pexels.com/video/video-of-travel-854669.