Converting instrument activations to segments?

[bmcfee]

I re-read the medleydb paper, and it doesn't actually say how the instrument segment annotations were computed from the activation functions. Is it just threshold at 0.5 and then run-length encode samples to intervals? Or is there some smoothing involved?

[rabitt]

Is it just threshold at 0.5 and then run-length encode samples to intervals?

Yes exactly. No smoothing.

[carlthome]

Rounding activations is often a lot worse than using some threshold on the first-order derivative (i.e. peak picking) to determine local max/min, no? Or this for example, is really good IMO (running averages).