deep neural networks are very successful on many vision tasks, but hard to
interpret due to their black box nature. To overcome this, various post-hoc
attribution methods have been proposed to identify image regions most
influential to the models' decisions. Evaluating such methods is