ImageNet-trained CNNs are biased towards texture; increasing shape bias\n improves accuracy and robustness

Convolutional Neural Networks (CNNs) are commonly thought to recognise\nobjects by learning increasingly complex representations of object shapes. Some\nrecent studies suggest a more important role of image textures. We here put\nthese conflicting hypotheses to a quantitative test by evaluating CNNs and\nhuman observers on images with a texture-shape cue conflict. We show that\nImageNet-trained CNNs are strongly biased towards recognising textures rather\nthan shapes, which is in stark contrast to human behavioural evidence and\nreveals fundamentally different classification strategies. We then demonstrate\nthat the same standard architecture (ResNet-50) that learns a texture-based\nrepresentation on ImageNet is able to learn a shape-based representation\ninstead when trained on "Stylized-ImageNet", a stylized version of ImageNet.\nThis provides a much better fit for human behavioural performance in our\nwell-controlled psychophysical lab setting (nine experiments totalling 48,560\npsychophysical trials across 97 observers) and comes with a number of\nunexpected emergent benefits such as improved object detection performance and\npreviously unseen robustness towards a wide range of image distortions,\nhighlighting advantages of a shape-based representation.\n