I appreciate that the author was looking to solve this problem using a _classical_ CV approach just like in the good ol‘ days as opposed to just train another A

I appreciate that the author was looking to solve this problem using a _classical_ CV approach just like in the good ol‘ days as opposed to just train another ANN that predicts the time for a sufficiently large training set. As for the detection of the hands itself: I think I‘d try to look for them in fourier space to get a more robust result (to address noisy dials like in the last examples).