Learn more about Israeli war crimes in Gaza, funded by the USA, Germany, the UK and others.

How to make a floating head

In a previous post. I showed how to run BodyPix on a video stream and access the segmentation from your shader. In this post, I demo the segmentPersonParts method, using it to make a floating head. You can run it on your own webcam;

A call to net.segmentPerson returns a Uint8Array where each value is either 1 (part of a person) or 0 (not part of a person). But the API also provides net.segmentPersonParts, which returns an Int32Array, but where the values are between -1 and 23, with -1 being not part of a person, and the other numbers being various parts. Here I’m just interested in values 0 and 1, which represent parts “left side of face” and “right side of face”.

To access this data in a WebGL shader, we need to get it into a texture using gl.texImage2D. When you pass an array to gl.texImage2D, you tell it which format to interpret it as. We’ll use gl.ALPHA, which has one unsigned byte per pixel. To convert to this format, we can use new Uint8Array(segmentation.data). The -1 values in signed ints get converted to 255 in unsigned ints. This byte interpreted as the alpha channel when the texture is accessed by a shader. Then in the pixel shader, these values between 0 and 255 are squeezed into the float range from 0 to 1. So we can check for the original values 1 and 2 with alpha <= 2./255..

BodyPix seems to include “neck” as part of the face. This is annoying for my purposes; I wanted to crop the face at the chin. I don’t see an easy way to hack around this.

Finally, here’s what I get when I run the demo against my own webcam feed:

Tagged #programming, #web, #webgl, #machine-learning.

Similar posts

More by Jim

Want to build a fantastic product using LLMs? I work at Granola where we're building the future IDE for knowledge work. Come and work with us! Read more or get in touch!

This page copyright James Fisher 2020. Content is not associated with my employer. Found an error? Edit this page.