To protect yourself during the outbreak of COVID-19, this webapp helps you keep your hands off your face using your webcam.
This webapp will issue a notification when it detects a hand is next to a face in the webcam feed. Running the application in background will help you catch and increase awareness of unintentional face touching.
We take your privacy very seriously. All of the camera data stays local and never leaves your computer.
An Apple Watch app is currently in development that uses the built-in sensors (accelerometers and gyroscope) to detect face touching so that you can increase your awareness when you are not in front of your computer. The app is featured in #BuildforCOVID19 Global Online Hackathon. We are hard at work in developing a beta version of the app. Sign up here if you want be our beta tester.
We utilized Google's handpose and facemesh models to detect when you're touching your face based on the intersection of your hand and face. Feel free to checkout our GitHub repository.
We are limited by the fact that the handpose model from Google can only recognize one hand.
Our goal is to make you aware of the unintentional face touches. It currently does not distinguish between other behaviors that occur around your face, such as drinking, touching glasses, putting in/taking out earbuds, or other activities with a close hand to face proximity. In addition, the models can sometimes classify a hand or a face incorrectly.
If the page is recognizing the touch correctly, make sure your notification settings is on.
If the page is not recognizing the touch correctly, there can be a few contributors. Since the app is relying on the camera, make sure your room has sufficient light and position yourself in the center of the camera. Your touch might be too fast for the system to register.
We are using the models trained by Google. It is possible that the data they used to train the models are biased and not a truly representative of all our use cases, and for different groups.
We are running two machine learning models in your browser, and that can consume quite some resources. Try increasing the frame timeout to reduce resource usage.
Currently, we've only tested on desktop environment. Due to the limitation of background tab on mobile browser, we are focusing our effort on desktop. Having said that, the basic recognition should still work when the browser window in in foreground on mobile browsers.
We'd like to hear about your experience of using the app. Tell us how we can improve, and what doesn't work for you here.