The matte computation has 2 stages. First is the binary matte computation, which is done in real time. In this stage each pixel is marked as either fully foreground or background.

In the second stage, which is done as an automatic post processing step, a soft/alpha matte for the pixels at the boundary of foreground and background pixels is computed. I agree that our second process is probably too heavy-handed in its feathering effect. Though these days there are better methods for computing alpha mattes for boundary pixels. For example - Anat Levin et al's "Closed Form Solution to Natural Image Matting."

Thats not very likely. The code is owned by Microsoft.

Aawa, what a shame. A project like this would be great in an open state. It could grow so fast... if only microsoft understood the concept of open source.

The skate boarder sequence shown in the video was 720x480 resolution and if I remember correctly a 150 frames long.

You would get higher quality mattes for higher resolution video. And you can always improve the quality of the matte by drawing more strokes.