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

Open paper

Benchmark	Model	Rank	Results
domain-generalization-on-imagenet-a	Stylized ImageNet (ResNet-50)	#37	Top-1 accuracy %: 2.3
domain-generalization-on-imagenet-c	Stylized ImageNet (ResNet-50)	#39	mean Corruption Error (mCE): 69.3
domain-generalization-on-imagenet-r	Stylized ImageNet (ResNet-50)	#33	Top-1 Error Rate: 58.5
domain-generalization-on-vizwiz	ResNet-50 (SIN_IN_IN)	#40	Accuracy - All Images: 39.2Accuracy - Corrupted Images: 32.4…
domain-generalization-on-vizwiz	ResNet-50 (SIN_IN)	#49	Accuracy - All Images: 38.2Accuracy - Corrupted Images: 32.5…
domain-generalization-on-vizwiz	ResNet-50 (SIN)	#86	Accuracy - All Images: 25.3Accuracy - Corrupted Images: 20.4…
object-recognition-on-shape-bias	AlexNet	#10	shape bias: 42.9
object-recognition-on-shape-bias	GoogLeNet	#14	shape bias: 31.2
object-recognition-on-shape-bias	ResNet-50	#16	shape bias: 22.1
object-recognition-on-shape-bias	VGG-16	#17	shape bias: 17.2