Can I use the sigmoid function to train a (n-1) classes segmentor in a (n) classes seg problem?

[WilliamJames0741]

There is a 5 classes seg problem, including the background. And I think the background is an important class in this problem, so I set the reduce_zero_label to False . Can I just train a 4 classes segmentor which use the sigmoid function and a threshold value to solve this 5 classes seg problem? What else do I need to modify, except the config file? Will its result be better than just train a simple 5-classes segmentor?