Use a virtual webcam background and overlays with body-pix and v4l2loopback
Virtual Webcam Background allows you to use a virtual background or to blur the background of your webcam image similar to commercial programs like Zoom.
Tensorflow with BodyPix is used to segment the image into foreground (person) and background using a neural network and v4l2loopback is used to create a virtual webcam.
As the script creates a virtual webcam device it works with any program that can use a v4l2 webcam.
See virtual-webcam.com for more information, image packs and more.
The program needs python 3.5–3.8 and is tested with python 3.7.
Install the requirements:
pip install -r requirements.txt
Download the bodypix model:
./get-model.sh
Then install v4l2loopback and load the kernel module:
modprobe v4l2loopback exclusive_caps=1 video_nr=2 # creates /dev/video2
The
exclusive_capsoption is needed by some programs, such as chromium.
Then copy
config.yaml.exampleto
config.yamland edit the config as needed and run the
virtual_webcam.pyscript.
If you have a Nvidia graphics card, you may want to install CUDA for better performance.
To configure the virtual webcam, edit
config.yaml. Most options are applied instantly, except for
widthand
heightas the webcam must be reinitialized to change them and
multiplierand
output_strideas the model must be reloaded to apply them.
width: The input resolution width.
height: The input resolution height.
fps: The input framerate.
mjpeg: Use mjpeg as input format. This may be faster than the default format.
segmentation_threshold: The threshold for foreground / background segmentation.
blur: Blur factor for the mask to smooth the edges.
dilate: Number of pixels the mask is shrunk to remove spots.
erode: Number of pixels the mask is grown after shrinking to capture the full body image again.
real_video_device: The video device of your webcam, e.g.
/dev/video0.
average_masks: Number of masks to average. A higher number will result in afterimages, a smaller number in flickering at the boundary between foreground and background.
layers: A list of videos layers like the input webcam image, the segmented foreground, virtual backgrounds or image overlays.
debug_show_mask: Debug option to show the mask, that can be used to configure blur/dilate/erode correctly.
model:
mobilenet(faster) or
resnet50(more accurate). You need to download the matching model, when you change the parameter.
multiplier: Multiplier parameter of the mobilenet model (0.5, 0.75 or 1.0). You need to download the matching model when you change this parameter.
output_stride: Stride parameter of the model (16 or 8 for
mobilenetand 16 or 32 for
resnet50). You need to download the matching model when you change the parameter.
internal_resolution: Resolution factor (between 0.0 and 1.0) for the model input. Smaller is faster and less accurate. Note that 1.0 does not always give the best results.
Note: Input
widthand
heightare autodetected when they are not set in the config, but this can lead to bad default values, e.g.,
640x480even when the camera supports a resolution of
1280x720.
The layers option contains one image source and a list of filters. The image sources are:
input: The webcam image
foreground: The foreground of the image, i.e., the person.
previous: The image composed of all previous layers.
empty: A transparent image.
Each layer has a list of filters, that are applied in the given order. After all filters are applied, the layer is merged with the previous layers.
Each layer has a list of filters. A simple example that converts the background to grayscale and blurs it looks like this:
- input: ["grayscale", "blur"]
Some filters have arguments. To change the blur value in the filter list above, you can use onf of these syntax variants:
["grayscale", ["blur", 10, 10]]
["grayscale", ["blur", [10, 10]]]
["grayscale", ["blur", {intensity_x: 10, intensity_y: 10}]]
- layers: - empty: [["image", "background.jpg"]] - foreground: []
- layers: - input: [["blur", 10]] - foreground: []
- layers: - input: [["blur", 10]] - foreground: [] - previous: [["image", "images/fog.jpg"], ["roll", 5, 0]]
The current filters and their options are:
image: Returns a static image, e.g., to use a virtual background.
image_path: The path to the image file.
interpolation_method: The interpolation method. Currently are
LINEARand
NEARESTsupported and
LINEARis the default. When you use a pixel art background, it may look better with
NEAREST.
image_sequence: Returns images from an image sequence. This can be used for animated backgrounds or overlays.
images_path: The path to a folder containing the images. The folder must only contain images and they must have the correct order when they are sorted by filename.
fps: The frames per second of the animation.
interpolation_method:
LINEARor
NEARESTinterpolation
video: Returns images from a video. This can be used for animated backgrounds or overlays.
video_path: The path to the video.
target_fps: The target frames per second of image sequence generated from the video. This can be used to reduce the RAM usage.
interpolation_method:
LINEARor
NEARESTinterpolation
blur: Blur the image.i
intensity_x: The intensity in the x direction.
intensity_y: The intensity in the y direction. When only
intensity_xis given, it will be used for
intensity_yas well.
gaussian_blur: Blur the image using a Gaussian blur. It looks better than normal box blur, but is more CPU intensive.
intensity_x: The intensity in the x direction. Must be an odd value: even values are bumped to the next odd value.
intensity_y: The intensity in the y direction. Must be an odd value: even values are bumped to the next odd value. When only
intensity_xis given, it will be used for
intensity_yas well.
grayscale: Convert the image into a grayscale image.
roll: move an image with a constant speed. This is mostly useful for overlays.
speed_x: Speed in x direction.
speed_y: Speed in y direction.
change_alpha: Change the transparency of an image.
change_alpha: Alpha value to add (between
-255and
255)
alpha_min,
alpha_max: Transparency levels to clip the resulting alpha value.
colorize: Change the image to grayscale and then color it with a given color.
r,
g,
b: RGB values.
color_filter: Change the color levels by multiplying the RGB values with a factor between
0and
255.
r,
g,
b: The factors for the colors red/green/blue.
solid_color: Fill the image with a single color.
r,
g,
b: RGB values.
flip: Flip the image horizontally or vertically.
horizontal: Flip horizontally.
vertical: Flip vertically.
stripes: Add a semi-transparent stripe effect with darker and lighter stripes.
width: Width of a stripe.
intensity: Intensity how much darker/lighter the stripe is.
speed: Speed at which the stripes move across the image.
chroma_key: Convert a color to transparency (green screen effect).
r,
g,
b: RGB values.
fuzz: Factor for fuzzy matching of similar colors.
If you have a video, you can use the
videofilter:
- "empty": [["video", "my-video.mp4"]]
Another option are image sequences, that allow for example to use transparent PNGs.
Example config for loading an image sequence from the folder "animation" and playing it with 5 frames per second:
- empty: [["image_sequence", "frames", 5]]
The program tries to load
frames/*.*and you need to make sure that the folder only contains images and that the images are ordered correctly when they are sorted by filename.
Example for creating an image sequence from a short video and adding alpha transparency for a green screen effect using ffmpeg and ImageMagick:
mkdir animation cd animation ffmpeg -i ../animation.webm -vf fps=10 out%04d.png mogrify -fuzz 10% -transparent 'rgb(0,129,27)' *
When using the
ffmpegcommand, you can change the output framerate using the
fpsparameter.
Note that the script loads all images of an animation into RAM scaled to the resolution of your webcam, so using too long animations is not a good idea.
To download other models get the full
get-model.shscript from https://github.com/ajaichemmanam/simple_bodypix_python and run it with one of these combinations:
./get-model.sh bodypix/mobilenet/float/{025,050,075,100}/model-stride{8,16} ./get-model.sh bodypix/resnet50/float/model-stride{16,32}
Example config for
mobilenet:
- model: mobilenet - multiplier: 0.5 - output_stride: 16
Example config for
resnet50:
- model: resnet50 - output_stride: 16
LICENSEfile.
bodypix_functions.pymay be used unter the terms of the MIT license.
bodypix_functions.pyare adapted from the body-pix nodejs module.