Which proves exactly what the top level comment claimed: the distribution of examples in the dataset was incorrect for the learning task.
Not necessarily. Textures are high frequency low abstraction data forms, shapes are low frequency high abstraction data forms.
The former are easier to learn for multiple reasons: The nature of higher frequency data means that you have more samples.
(There are far more furry textures in a picture of a cat than there are toes, paws, legs etc.)
Low abstraction data forms are also learnt in earlier layers of the network - more abstract forms are resolved through the understanding of how less abstract forms group together, making the low abstraction forms a prerequisite to more in depth learning.
But the insight is that although both features are informative, classical training approaches result in networks that over-fit on texture information and ignore shape information.
Table 2 of the discussed paper finds that joint training with "Stylized ImageNet" (IN images after a style transfer) and classic IN followed by a fine-tuning pass on IN results in improved classification accuracy over IN training alone.