A control dataset that tests whether an image classifier relies on global object shape or local texture, following Geirhos et al., ImageNet-trained CNNs are biased towards texture (ICLR 2019, arXiv:1811.12231).
control-dataset/— generator, evaluator, and the dataset (data/original,data/edges,data/texture_patch). See its README to run.report/— the blog post (blog.md).
A standard ResNet-50 scores 99.5% on the original images, 98.0% when the shape is scrambled, and 26.0% when only the shape (edges) remains: a +72 point texture bias.