We do not recompute probabilities when the frame and the ROI do
not change. It saves a lot of time when adjusting the probability
threshold, because ilastik takes a lot of time to launch everytime.
Here we used the same strategy that for the Weka detector: we
simply store the last probability map and check whether we need
to recompute it next time the user presses the 'Preview' button.
Nowe I have to check that it also works for 3D.

…red.

Amazingly enough, it is sufficient to simply add the smoothing
parameter to the settings maps for it to work. Indeed, the smoothing
is handled by the threshold detector, that for ilastik, operates
on the probability map generated by the ilastik command. So there
is only very little to do to make it work here.
Unexpected but pleasant.