How to train a multi-objective semantic segmentation data set

[Atumyugi]

Different from the example data set, each image in my data set has multiple targets that need to be identified. The categories of these targets may not be consistent. Please follow the example and change the multiple targets to 0 255. Can it be trained? I only need the identification ability of SAM and do not need the classification ability. Or do I need to change the data set to a data set with only one target unit?
eg：

i need identifiy multi building

[Atumyugi]

If i follow the train example and change the multiple targets to single 0 255. Can it be trained?

[WuJunde]

We've supported multi-class segmentation (maximum 4 classes). Use it by setting -multimask_output to the number of classes favored. Also updated REFUGE example to two classes (optic disc & cup). Credit: @LJQCN101