abstraction, 1p, The results are no worse than their ImageNet pre-training counterparts even when using the hyper-parameters of the baseline system (심지어 mask-CNN 은 fine tuning 에 최적화된 system 임에도 불구하고)