Ignoring background as class

[AKMourato]

I notice that the current structure provided reads labels with grayscale annotations in png format and count the background as a class. Can this be avoided? Assigning 255 to background pixels in order to be excluded for training is an option?