Does DetectNet generate a ground truth coverage map for each class?

[JustinMBrown]

When doing multiclass object detection training, does DetectNet generate a single ground truth coverage map such that each grid space can have 1 out of N classes, or does it generate a separate ground truth coverage map for each N classes? Because the output of a multiclass DetectNet is definitely a separate coverage map for each class, but I'm not convinced the ground truth is.

And can you point me to the piece of code where this is happening?