Linux:movie-camera:

April 9th, 2020


With many of us around the globe under shelter in place due to COVID-19
video calls have become a lot more common. In particular, ZOOM has
controversially become very popular. Arguably Zoom’s most interesting feature
is the “Virtual Background” support which allows users to replace
the background behind them in their webcam video feed with any image (or video).

I’ve been using Zoom for a long time at work for Kubernetes open source meetings,
usually from my company laptop. With daily “work from home” I’m now inclined to
use my more powerful and ergonomic personal desktop for some of my open source work.

Unfortunately, Zoom’s linux client only supports the “chroma-key” A.K.A. “green screen
background removal method. This method requires a solid color backdrop, ideally
a green screen with uniform lighting.

Since I do not have a green screen I decided to simply implement my own background
removal, which was obviously better than cleaning my apartment or just using
my laptop all the time. 😀

It turns out we can actually get pretty decent results with off the shelf, open source
components and just a little of our own code.

Reading The Camera 🔗︎

First thing’s first: How are we going to get the video feed from our webcam for
processing?

Since I use Linux on my personal desktop (when not playing PC games) I chose to
use the OpenCV python bindings as I’m already familiar with them and they include
useful image processing primatives in addition to V4L2 bindings for reading from
webcams.

Reading a frame from the webcam with python-opencv is very simple:

1import cv2
2cap = cv2.VideoCapture('/dev/video0')
3success, frame = cap.read()

For better results with my camera before capturing set:

1# configure camera for 720p @ 60 FPS
2height, width = 720, 1280
3cap.set(cv2.CAP_PROP_FRAME_WIDTH ,width)
4cap.set(cv2.CAP_PROP_FRAME_HEIGHT,height)
5cap.set(cv2.CAP_PROP_FPS, 60)

Most video conferencing software seems to cap video to 720p @ 30 FPS or lower,
but we won’t necessarily read every frame anyhow, this sets an upper limit.

Put the frame capture in a loop and we’ve got our video feed!

1while True:
2    success, frame = cap.read()

We can save a test frame with just:

1cv2.imwrite("test.jpg", frame)

And now we can see that our camera works. Success!

don’t mind my corona beard

don’t mind my corona beard

Finding The Background 🔗︎

OK, now that we have a video feed, how do we identify the background so we can
replace it? This is the tricky part …

While Zoom doesn’t seem to have commented anywhere about how they implemented
this, the way it behaves makes me suspect that a neural network is involved,
it’s hard to explain but the results look like one.
Additionally, I found an article about Microsoft Teams implementing background blur with a convolutional neural network.

Creating our own network wouldn’t be too hard in principle – There are many
articles and papers on the topic of image segmentation and plenty of open
source libraries and tools, but we need a fairly specialized dataset to get
good results.

Specifically we’d need lots of webcam like images with the ideal
human foreground marked pixel by pixel versus the background.

Building this sort of dataset in prepartion for training a neural net probably would
be a lot of work. Thankfully a research team at Google has already done all of this hard
work and open sourced a pre-trained neural network for “person segmentation”
called BodyPix that works pretty well! ❤️

BodyPix is currently only available in TensorFlow.js form, so the easiest
way to use it is from the body-pix-node library.

To get faster inference (prediction) in the browser a WebGL backend is preferred, but in
node we can use the Tensorflow GPU backend
(NOTE: this requires a NVIDIA Graphics Card, which I have).

To make this easier to setup, we’ll start by setting up a small containerized
tensorflow-gpu + node environment / project. Using this with nvidia-docker is
much easier than getting all of the right dependencies setup on your host, it
only requires docker and an up-to-date GPU driver on the host.

1{
2    "name": "bodypix",
3    "version": "0.0.1",
4    "dependencies": {
5        "@tensorflow-models/body-pix": "^2.0.5",
6        "@tensorflow/tfjs-node-gpu": "^1.7.1"
7    }
8}

bodypix/DockerfileDockerfile

 1# Base image with TensorFlow GPU requirements
 2FROM nvcr.io/nvidia/cuda:10.0-cudnn7-runtime-ubuntu18.04
 3# Install node
 4RUN apt update && apt install -y curl make build-essential 
 5    && curl -sL https://deb.nodesource.com/setup_12.x | bash - 
 6    && apt-get -y install nodejs 
 7    && mkdir /.npm 
 8    && chmod 777 /.npm
 9# Ensure we can get enough GPU memory
10# Unfortunately tfjs-node-gpu exposes no gpu configuration :(
11ENV TF_FORCE_GPU_ALLOW_GROWTH=true
12# Install node package dependencies
13WORKDIR /src
14COPY package.json /src/
15RUN npm install
16# Setup our app as the entrypoint
17COPY app.js /src/
18ENTRYPOINT node /src/app.js

Now to serve the results… WARNING: I am not a node expert! This is just
my quick evening hack, bear with me 🙂

The following simple script replies to an HTTP POSTed image with a binary mask
(an 2d array of binary pixels, where zero pixels are the background).

 1const tf = require('@tensorflow/tfjs-node-gpu');
 2const bodyPix = require('@tensorflow-models/body-pix');
 3const http = require('http');
 4(async () => {
 5    const net = await bodyPix.load({
 6        architecture: 'MobileNetV1',
 7        outputStride: 16,
 8        multiplier: 0.75,
 9        quantBytes: 2,
10    });
11    const server = http.createServer();
12    server.on('request', async (req, res) => {
13        var chunks = [];
14        req.on('data', (chunk) => {
15            chunks.push(chunk);
16        });
17        req.on('end', async () => {
18            const image = tf.node.decodeImage(Buffer.concat(chunks));
19            segmentation = await net.segmentPerson(image, {
20                flipHorizontal: false,
21                internalResolution: 'medium',
22                segmentationThreshold: 0.7,
23            });
24            res.writeHead(200, { 'Content-Type': 'application/octet-stream' });
25            res.write(Buffer.from(segmentation.data));
26            res.end();
27            tf.dispose(image);
28        });
29    });
30    server.listen(9000);
31})();

We can use numpy and requests to convert a frame to a mask from our
python script with the following method:

 1def get_mask(frame, bodypix_url='http://localhost:9000'):
 2    _, data = cv2.imencode(".jpg", frame)
 3    r = requests.post(
 4        url=bodypix_url,
 5        data=data.tobytes(),
 6        headers={'Content-Type': 'application/octet-stream'})
 7    # convert raw bytes to a numpy array
 8    # raw data is uint8[width * height] with value 0 or 1
 9    mask = np.frombuffer(r.content, dtype=np.uint8)
10    mask = mask.reshape((frame.shape[0], frame.shape[1]))
11    return mask

Which gives us a result something like:

Open-Source Virtual Background 1

While I was working on this, I spotted this tweet:

Now that we have the foreground / background mask, it will be easy to replace
the background.

After grabbing the awesome “Virtual Background” picture from that twitter thread and
cropping it to a 16:9 ratio image …

Open-Source Virtual Background 3

… we can do the following:

 1# read in a "virtual background" (should be in 16:9 ratio)
 2replacement_bg_raw = cv2.imread("background.jpg")
 3
 4# resize to match the frame (width & height from before)
 5width, height = 720, 1280
 6replacement_bg = cv2.resize(replacement_bg_raw, (width, height))
 7
 8# combine the background and foreground, using the mask and its inverse
 9inv_mask = 1-mask
10for c in range(frame.shape[2]):
11    frame[:,:,c] = frame[:,:,c]*mask + replacement_bg[:,:,c]*inv_mask

Which gives us:

Open-Source Virtual Background 4

The raw mask is clearly not tight enough due to the performance trade-offs
we made with our BodyPix parameters but .. so far so good!

This background gave me an idea …

Now that we have the masking done, what can we do to make it look better?

The first obvious step is to smooth the mask out, with something like:

1def post_process_mask(mask):
2    mask = cv2.dilate(mask, np.ones((10,10), np.uint8) , iterations=1)
3    mask = cv2.erode(mask, np.ones((10,10), np.uint8) , iterations=1)
4    return mask

This can help a bit, but it’s pretty minor and just replacing the background
is a little boring, since we’ve hacked this up ourselves we can do anything
instead of just a basic background removal …

Given that we’re using a Star Wars “virtual background” I decided to create
hologram effect to fit in better. This also lets lean into blurring the mask.

First update the post processing to:

1def post_process_mask(mask):
2    mask = cv2.dilate(mask, np.ones((10,10), np.uint8) , iterations=1)
3    mask = cv2.blur(mask.astype(float), (30,30))
4    return mask

Now the edges are blurry which is good, but we need to start building the hologram
effect.

Hollywood holograms typically have the following properties:

  • washed out / monocrhomatic color, as if done with a bright laser
  • scan lines or a grid like effect, as if many beams created the image
  • “ghosting” as if the projection is done in layers or imperfectly reaching the correct distance

We can add these step by step.

First for the blue tint we just need to apply an OpenCV colormap:

1# map the frame into a blue-green colorspace
2holo = cv2.applyColorMap(frame, cv2.COLORMAP_WINTER)

Then we can add the scan lines with a halftone-like effect:

1# for every bandLength rows darken to 10-30% brightness,
2# then don't touch for bandGap rows.
3bandLength, bandGap = 2, 3
4for y in range(holo.shape[0]):
5    if y % (bandLength+bandGap) < bandLength:
6        holo[y,:,:] = holo[y,:,:] * np.random.uniform(0.1, 0.3)

Next we can add some ghosting by adding weighted copies of the current effect,
shifted along an axis:

 1# shift_img from: https://stackoverflow.com/a/53140617
 2def shift_img(img, dx, dy):
 3    img = np.roll(img, dy, axis=0)
 4    img = np.roll(img, dx, axis=1)
 5    if dy>0:
 6        img[:dy, :] = 0
 7    elif dy<0:
 8        img[dy:, :] = 0
 9    if dx>0:
10        img[:, :dx] = 0
11    elif dx<0:
12        img[:, dx:] = 0
13    return img
14
15# the first one is roughly: holo * 0.2 + shifted_holo * 0.8 + 0
16holo2 = cv2.addWeighted(holo, 0.2, shift_img(holo1.copy(), 5, 5), 0.8, 0)
17holo2 = cv2.addWeighted(holo2, 0.4, shift_img(holo1.copy(), -5, -5), 0.6, 0)

Last: We’ll want to keep some of the original color, so let’s combine
the holo effect with the original frame similar to how we added the ghosting:

1holo_done = cv2.addWeighted(img, 0.5, holo2, 0.6, 0)

A frame with the hologram effect now looks like:

Open-Source Virtual Background 5

On it’s own this looks pretty :shrug:

But combined with our virtual background it looks more like:

Open-Source Virtual Background 6

There we go! :tada: (I promise it looks cooler with motion / video :upside_down_face:)

Outputting Video 🔗︎

Now we’re just missing one thing … We can’t actually use this in a call yet.

To fix that, we’re going to use pyfakewebcam and v4l2loopback to create a fake webcam device.

We’re also going to actually wire this all up with docker.

First create a requirements.txt with our dependencies:

fakecam/requirements.txtDockerfile

1numpy==1.18.2
2opencv-python==4.2.0.32
3requests==2.23.0
4pyfakewebcam==0.1.0

And then the Dockerfile for the fake camera app:

fakecam/DockerfileDockerfile

 1FROM python:3-buster
 2# ensure pip is up to date
 3RUN pip install --upgrade pip
 4# install opencv dependencies
 5RUN apt-get update && 
 6    apt-get install -y 
 7      `# opencv requirements` 
 8      libsm6 libxext6 libxrender-dev 
 9      `# opencv video opening requirements` 
10      libv4l-dev
11# install our requirements
12WORKDIR /src
13COPY requirements.txt /src/
14RUN pip install --no-cache-dir -r /src/requirements.txt
15# copy in the virtual background
16COPY background.jpg /data/
17# run our fake camera script (with unbuffered output for easier debug)
18COPY fake.py /src/
19ENTRYPOINT python -u fake.py

We’re going to need to install v4l2loopback from a shell:

1sudo apt install v4l2loopback-dkms

And then configure a fake camera device:

1sudo modprobe -r v4l2loopback
2sudo modprobe v4l2loopback devices=1 video_nr=20 card_label="v4l2loopback" exclusive_caps=1

We need the exclusive_caps setting for some apps (chrome, zoom) to work, the label
is just for our convenience when selecting the camera in apps, and the video number
just makes this /dev/video20 if available, which is unlikely to be already in use.

Now we can update our script to create the fake camera:

1# again use width, height from before
2fake = pyfakewebcam.FakeWebcam('/dev/video20', width, height)

We also need to note that pyfakewebcam expects images in RGB (red, green, blue)
while our OpenCV operations are in BGR (blue, green, red) channel order.

We can fix this before outputting and then send a frame with:

1frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
2fake.schedule_frame(frame)

All together the script looks like:

 1import os
 2import cv2
 3import numpy as np
 4import requests
 5import pyfakewebcam
 6
 7def get_mask(frame, bodypix_url='http://localhost:9000'):
 8    _, data = cv2.imencode(".jpg", frame)
 9    r = requests.post(
10        url=bodypix_url,
11        data=data.tobytes(),
12        headers={'Content-Type': 'application/octet-stream'})
13    mask = np.frombuffer(r.content, dtype=np.uint8)
14    mask = mask.reshape((frame.shape[0], frame.shape[1]))
15    return mask
16
17def post_process_mask(mask):
18    mask = cv2.dilate(mask, np.ones((10,10), np.uint8) , iterations=1)
19    mask = cv2.blur(mask.astype(float), (30,30))
20    return mask
21
22def shift_img(img, dx, dy):
23    img = np.roll(img, dy, axis=0)
24    img = np.roll(img, dx, axis=1)
25    if dy>0:
26        img[:dy, :] = 0
27    elif dy<0:
28        img[dy:, :] = 0
29    if dx>0:
30        img[:, :dx] = 0
31    elif dx<0:
32        img[:, dx:] = 0
33    return img
34
35def hologram_effect(img):
36    # add a blue tint
37    holo = cv2.applyColorMap(img, cv2.COLORMAP_WINTER)
38    # add a halftone effect
39    bandLength, bandGap = 2, 3
40    for y in range(holo.shape[0]):
41        if y % (bandLength+bandGap) < bandLength:
42            holo[y,:,:] = holo[y,:,:] * np.random.uniform(0.1, 0.3)
43    # add some ghosting
44    holo_blur = cv2.addWeighted(holo, 0.2, shift_image(holo.copy(), 5, 5), 0.8, 0)
45    holo_blur = cv2.addWeighted(holo_blur, 0.4, shift_image(holo.copy(), -5, -5), 0.6, 0)
46    # combine with the original color, oversaturated
47    out = cv2.addWeighted(img, 0.5, holo_blur, 0.6, 0)
48    return out
49
50def get_frame(cap, background_scaled):
51    _, frame = cap.read()
52    # fetch the mask with retries (the app needs to warmup and we're lazy)
53    # e v e n t u a l l y c o n s i s t e n t
54    mask = None
55    while mask is None:
56        try:
57            mask = get_mask(frame)
58        except:
59            print("mask request failed, retrying")
60    # post-process mask and frame
61    mask = post_process_mask(mask)
62    frame = hologram_effect(frame)
63    # composite the foreground and background
64    inv_mask = 1-mask
65    for c in range(frame.shape[2]):
66        frame[:,:,c] = frame[:,:,c]*mask + background_scaled[:,:,c]*inv_mask
67    return frame
68
69# setup access to the *real* webcam
70cap = cv2.VideoCapture('/dev/video0')
71height, width = 720, 1280
72cap.set(cv2.CAP_PROP_FRAME_WIDTH, width)
73cap.set(cv2.CAP_PROP_FRAME_HEIGHT, height)
74cap.set(cv2.CAP_PROP_FPS, 60)
75
76# setup the fake camera
77fake = pyfakewebcam.FakeWebcam('/dev/video20', width, height)
78
79# load the virtual background
80background = cv2.imread("/data/background.jpg")
81background_scaled = cv2.resize(background, (width, height))
82
83# frames forever
84while True:
85    frame = get_frame(cap, background_scaled)
86    # fake webcam expects RGB
87    frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
88    fake.schedule_frame(frame)

Now build the images:

1docker build -t bodypix ./bodypix
2docker build -t fakecam ./fakecam

And run them like:

 1# create a network
 2docker network create --driver bridge fakecam
 3# start the bodypix app
 4docker run -d 
 5  --name=bodypix 
 6  --network=fakecam 
 7  --gpus=all --shm-size=1g --ulimit memlock=-1 --ulimit stack=67108864 
 8  bodypix
 9# start the camera, note that we need to pass through video devices,
10# and we want our user ID and group to have permission to them
11# you may need to `sudo groupadd $USER video`
12docker run -d 
13  --name=fakecam 
14  --network=fakecam 
15  -p 8080:8080 
16  -u "$$(id -u):$$(getent group video | cut -d: -f3)" 
17  $$(find /dev -name 'video*' -printf "--device %p ") 
18  fakecam

Now make sure to start this before opening the camera with any apps, and
be sure to select the “v4l2loopback” / /dev/video20 camera in Zoom etc.

The Finished Result 🔗︎

Here’s a quick clip I recorded of this in action:

Look! I’m dialing into the millenium falcon with an open source camera stack!

I’m pretty happy with how this came out. I’ll definitely be joining all of my meetings this way in the morning. 😀

LEAVE A REPLY

Please enter your comment!
Please enter your name here