MSRC ground truth comes as RGB images. We turn these into integer label images. Therefore the labels are defined by us. They have been chosen arbitrarily. Currently "void" is the first listed class so it has the label 0. This throws everything into disarray because during classification void is not included. At various points the labels are shifted by 1 to cope with this, which is hell confusing.
Elegant solution: move void to the end of the class list. It is then the last label, and the subset that excludes it is not affected by it.
In the future it would be more prudent to make it something ridiculously big so it's obviously invalid.
MSRC ground truth comes as RGB images. We turn these into integer label images. Therefore the labels are defined by us. They have been chosen arbitrarily. Currently "void" is the first listed class so it has the label 0. This throws everything into disarray because during classification void is not included. At various points the labels are shifted by 1 to cope with this, which is hell confusing.
Elegant solution: move void to the end of the class list. It is then the last label, and the subset that excludes it is not affected by it.
In the future it would be more prudent to make it something ridiculously big so it's obviously invalid.