Discord for Linux currently doesn’t support background blurring, which is a good excuse for a side project.

So, after diving into a rabbit hole filled with OpenCV and virtual devices I’ve emerged with a solution.

CustomCam is a Python framework that makes it super easy to build and apply your own video effects to a live webcam.

Try it out now with pip install CustomCam.

Filter-First Framework

Most OpenCV-based blogs & tutorials talk a lot about webcam capture and image manipulation but don’t concern themselves with outputting those modified frames elsewhere (i.e. to Discord or Zoom). CustomCam aims to fill this gap by implementing a framework that allows users to focus on building filters not infrastructure.

You can think of the CustomCam pipeline as the following three components:

1. Input: Frames are received from your regular webcam
2. Filter: Video effects are applied to each frame
   • The active filter can be changed mid-stream from the command line
3. Output: Modified frames are passed to a virtual output device
   • You can create these with v4l2loopback or OBS
   • More information can be found in the pyvirtualcam docs

Filter Buffet

Filter	Effect	Notes
NoFilter	Removes filter
Gray	Super simple grayscale
Sepia	Standard sepia effect	Based on a gist by FilipeChagasDev
Segment	Silhouette-based background-blurring	Utilises MediaPipe’s Selfie Segmentation. Definitely the best approach for background-blurring.
Pixel	Silhouette-based foreground-blurring	Utilises MediaPipe’s Selfie Segmentation. Essentially an inversion of Segment.
BlurSat	Saturation-based background blurring	Based on blog by PythonWithRune. I wasn’t able to effectively optimise this for a diversity of frames.
BlurBox	Facial recognition-based background blurring	Based on blog by data-stats. This works fairly well, but the facial detection is a bit rough. Rotating your head breaks it.
Frame	Facial recognition-based	Essentially a debug version of BlurBox. Adds a green box around faces.

Creating your own filters

CustomCam is built to be customisable, so it’s super simple to add your own.

Simply create a new class in filters.py that:

• Inherits from the filters.Filter class
• Implements a __str__ method, which returns a string containing a short description of the filter.
• Implements an apply method, which applies your filter logic to each frame and returns a np.array.
• Doesn’t share a name with any existing class or input command.

Filter class names are automatically detected on launch, the class name will become its run command.