The voronoi is just used to map the x,y coordinates to tiles that then map to images. Doesn't say how the images were selected and the finger coordinates mapped in the first place to generate that map.
I could be wrong but when this was created, pre-2012, it would have been quite hard to get a computer to annotate those pictures accurately. 2012 was the year that kick-started the deep learning revolution[0], which greatly simplified the realization of such tasks.
Reminds me of a recent HN comment about the difference between mechanical and mathematical magic tricks.
"A magician I know talks about the three kinds of magic: sleight of hand, mechanical, and mathematical. The first requires a lot of skill and practice (e.g., palming coins, manipulating cards). The second requires a lot of preparation and engineering (e.g., sawing a person in half, floating person illusions). The third relies on the nature of reality (e.g., dividing sets of cards in a known pattern that forces a result, or manipulating numbers that force an unexpected result)."
They do not, at least in this implementation they do not. If you watch the video, pointing finger position is actually preset in the data (by human analysis most likely). Voronoi (it is pronounced with “oy” not “ay” btw) is only used for proximity matching.
https://youtube.com/watch?v=Z2ZXW2HBLPM