README.md

Google MediaPipe Demo by Joseph Writes Code

This is a simple demo of Google's MediaPipe library using TypeScript and React.

The implementation showcases the Pose Landmarker and Face Landmarker detection models in real-time using the webcam.

Google MediaPipe is a powerful library for real-time machine learning inference. It's particularly useful for pose and face detection, and it's available as a WebAssembly module.

The underlying models are based on TensorFlow Lite, and the library is designed to be fast and efficient. It's a great choice for real-time applications, and it's relatively easy to use in the browser.

Who is this for?

• Web developers who are curious about using MediaPipe in their projects.

💬 Try chatting with a custom GPT about this repo!

Installation

1. Clone this repository
2. Run npm install to install dependencies
3. The post-install script will automatically download the necessary MediaPipe models and wasm dependencies, and place them in the public directory.

Usage

1. Run npm run storybook to start the Storybook server.

Implemented Features

lib/MediaPipe.ts

• This file offers a light abstraction layer for the MediaPipe library. Rather than using the MediaPipe JavaScript API directly, you can call getVideoPoseLandmarker and getVideoFaceLandmarker to instantiate the respective PoseLandmarker and FaceLandmarker models from MediaPipe.

• It also offers a few smoothing functions to help stabilize the landmark points.

  • getLandmarkSmootherEWMA uses an Exponential Weighted Moving Average to smooth the landmark points.

  • getLandmarkSmootherKalman uses a Kalman filter to smooth the landmark points. (This is implemented in lib/KalmanFilter.ts) You may find that this is more effective than the EWMA smoother, since you can tune the algorithm to the expected noise in the data.

components/<WebcamLandmarked>

• This is a React component that uses the lib/MediaPipe.ts abstraction to display the webcam feed with the landmark points overlaid.

• It also offers a few props to control the landmark smoothing, or really perform whatever post-processing you would like on the landmark points.

  import {createWebcamLandmarked} from "./WebcamLandmarked.tsx";
  import {getLandmarkSmootherEWMA} from "@/lib/MediaPipe.ts";
  
  const WebcamLandmarked = createWebcamLandmarked();
  
  <WebcamLandmarked
  	getLandmarkPostProcessingFn={() => getLandmarkSmootherEWMA({strength: 0.5})}
  />;

• ⚠️ Note: The device delegation for inference is hard-coded to GPU.

  • You may want to change this to CPU if you're running on a low-powered device. This can be done by modifying the initPoseLandmarker method in components/WebcamLandmarked.tsx as follows:

    useEffect(() => {
    	(async function initPoseLandmarker() {
    		poseLandmarkerRef.current = await getVideoPoseLandmarker({
    			// Change to MEDIAPIPE.DELEGATE.CPU as needed.
    			delegate: MEDIAPIPE.DELEGATE.GPU,
    		});
    		faceLandmarkerRef.current = await getVideoFaceLandmarker({
    			// Change to MEDIAPIPE.DELEGATE.CPU as needed.
    			delegate: MEDIAPIPE.DELEGATE.GPU,
    		});
    	})();
    }, []);

    You could also modify this to be a prop of the component, though you will likely need to modify useEffect to handle the change in delegate.

• Similarly, the default model used for Pose Estimation is always set to "HEAVY".

  • You can change this by also modifying the initPoseLandmarker method in components/WebcamLandmarked.tsx as follows:

    useEffect(() => {
    	(async function initPoseLandmarker() {
    		poseLandmarkerRef.current = await getVideoPoseLandmarker({
    			// Change to "LITE", "HEAVY", or "FULL" as needed.
    			modelType: "HEAVY",
    			delegate: MEDIAPIPE.DELEGATE.GPU,
    		});
    		faceLandmarkerRef.current = await getVideoFaceLandmarker({
    			// There is no model option for Face Landmarking.
    			delegate: MEDIAPIPE.DELEGATE.GPU,
    		});
    	})();
    }, []);

components/<Webcam>

• This is a simple React component that uses the getUserMedia API to display the webcam feed. It's used by <WebcamLandmarked>. You can use it as a standalone component if you need to display the webcam feed without any landmark detection.

License

The demo is available as open source under the terms of the MIT License.