The class 0 of semantic labels

[sysu19351158]

Thanks for your excellent work. I read the code carefully. And I notice that when calculating the CE loss and the miou, the class 0 seems to be ignored, why?

[danielS91]

Class 0 is the void class that indicates an invalid pixel or a missing annotation (for the datasets we consider). Thus, pixels of the void class should be ignored during loss and metric calculation.
Note that our semantic head also does not predict this class, it can only occur after panoptic merging in predictions (if --instance-offset-distance-threshold is used).

[sysu19351158]

Thanks a lot! I understand now.