Table 1 Shape bias and accuracy of the...

Table 1

Shape bias and accuracy of the ImageNet validation set of ResNet50. IN: ImageNet, SIN: Stylized-ImageNet, SmIN: Smooth-ImageNet (proposed method), L0IN: L0-ImageNet (proposed method).

pretraining data	Fine-tuning data	Shape bias	TOP-1	TOP-5
IN	-	0.214	76.1	92.0
SIN + IN	-	0.370	64.6	85.1
SIN + IN	IN	0.362	74.1	91.9
SmIN + IN	-	0.266	75.4	92.5
SmIN + IN	IN	0.267	76.8	93.2
LOIN + IN	-	0.269	72.2	90.6
LOIN + IN	IN	0.272	76.3	93.1

pretraining data	Fine-tuning data	Shape bias	TOP-1	TOP-5
IN	-	0.214	76.1	92.0
SIN + IN	-	0.370	64.6	85.1
SIN + IN	IN	0.362	74.1	91.9
SmIN + IN	-	0.266	75.4	92.5
SmIN + IN	IN	0.267	76.8	93.2
LOIN + IN	-	0.269	72.2	90.6
LOIN + IN	IN	0.272	76.3	93.1

[ViewLarge]

Sharing Unavailable