shape bias

The 'shape bias' dataset was introduced in Geirhos et al. (ICLR 2019) and consists of 224x224 images with conflicting texture and shape information (e.g., cat shape with elephant texture). This is used to measure the shape vs. texture bias of image classifiers.