I will say, though, that I hate the mouse UI: you can't use relative offsets to make changes on a device that might disappear form one edge of your control and reappear on any other. It's very confusing to the user. Please use a drag for that.
That got me wondering about what sort of tricks might be out there.
For human users, you could use a face detection algorithm to find faces in the image and then place smoothed face meshes (or even just blobs) under them, and you'd have a fake depth-camera that would be pretty convincing until someone put their hand up.
Googling around, I found that there's a lot of research being done in depth inference from video. This paper: http://people.csail.mit.edu/celiu/pdfs/ECCV12-autostereo.pdf is particularly impressive, although given their one minute/frame benchmarks, we'll be waiting 15 years or so before Moore's law can bring this technique to a JS demo.
A demo: http://vimeo.com/8297993
Source code: https://github.com/hcarvalhoalves/processing-cammesh
Nice hack though. Can't wait to see some real CV with WebGL.
It was interesting using Manycam to use the area around my mouse pointer as the video source and let it do the effect to itself recursively.