This is an app to demonstrate the use of VisionCamera to build a realtime face blurring application. π·
RPReplay_Final1713800902.MP4
It took me 25 minutes to build the app, and this runs on both iOS and Android at 60-120 FPS. I can swap out the drawing algorithm, plug in a different shader, use a different ML model like pose-, object- or hand-detection, and change Camera properties within seconds to minutes - it has never been so easy to build Camera apps, all thanks to react-native-vision-camera! π
FaceBlurApp uses react-native-vision-camera to display a Camera feed.
Using the VisionCamera Frame Processors API and the react-native-vision-camera-face-detector Frame Processor plugin, we can detect faces in realtime at 60-120 FPS.
Then, we can draw a blur shader/mask over the detected faces by using the VisionCamera Skia Frame Processors integration. In this case, a simple blur ImageFilter
from react-native-skia is used.
This is the relevant code:
const {detectFaces} = useFaceDetector({
performanceMode: 'fast',
contourMode: 'all',
landmarkMode: 'none',
classificationMode: 'none',
});
const blurRadius = 25;
const blurFilter = Skia.ImageFilter.MakeBlur(
blurRadius,
blurRadius,
TileMode.Repeat,
null,
);
const paint = Skia.Paint();
paint.setImageFilter(blurFilter);
const frameProcessor = useSkiaFrameProcessor(frame => {
'worklet';
// 1. Render frame as it is
frame.render();
// 2. Detect faces in frame
const {faces} = detectFaces({frame: frame});
// 3. Draw a blur mask over each face
for (const face of faces) {
const path = Skia.Path.Make();
const necessaryContours: (keyof Contours)[] = [
'FACE',
'LEFT_CHEEK',
'RIGHT_CHEEK',
];
for (const key of necessaryContours) {
const points = face.contours[key];
points.forEach((point, index) => {
if (index === 0) {
// it's a starting point
path.moveTo(point.x, point.y);
} else {
// it's a continuation
path.lineTo(point.x, point.y);
}
});
path.close();
}
frame.save();
frame.clipPath(path, ClipOp.Intersect, true);
frame.render(paint);
frame.restore();
}
}, [paint, detectFaces])
See
App.tsx
for the full code.
If you know me you know that my libs are always really fast. This is no exception:
- Frame Processors are built almost entirely with C++ to run Frame analysis with almost zero overhead.
- Frame Processors run in Worklets on a separate Thread, so they are not interrupted by JS lags.
- The MLKit models are built with native code and use GPU-acceleration
- VisionCamera can provide either
yuv
or efficiently convertedrgb
buffers for faster ML execution - VisionCamera supports GPU buffer compression