-
The segmentation is done on 4 sets of data. One segmentation map per RGB channel and one for “all blown”
-
The color in the combining mask is a false-color representation of the segments found for each color channel. What you can see is the red channel has not been connected to other reds.
-
color in canditation mask, again a false color representation of segments that got a “good candidate”. Unfortunately the algorithm for selecting the best-candidate couldn’t find a good-enough one for the reds even at 100%. So failing here
And yes, selecting combine and candidating is trial&error 
But you should remember in which situations the algorithm might be a good one to try.