semantic segmentation is still a challenging task for parsing diverse
contexts in different scenes, thus the fixed classifier might not be able to
well address varying feature distributions during testing. Different from the
mainstream literature where the efficacy of strong backbones