Supplementary Materials

Figure S1: Class distribution of the IN118K-LT dataset. We construct the IN-118K dataset by sampling a uniform subset from the ImageNet train split. Differently, the IN-118KLT is a long-tailed version of the ImageNet train split. The classes follow the Pareto distribution with power value α = 6 (see Figure S1). OI-118K is obtained by randomly sampling 118K images from the OpenImages-v4 train split. Note that all datasets are constructed to be of similar size, which facilitates an apples-to-apples comparison.