The Origins and Prevalence of Texture Bias in Convolutional Neural Networks

However, all objectives and architectures still lead to models that make texture-based classification decisions a majority of the time, even if shape information is decodable from their hidden representations. The effect of data augmentation is much larger.