Table 3.

Test accuracy comparison between original ResNets and ResNets with proposed SR under incremental augmentation levels. ResNet is abbreviated as R. The scale factor scheduler for scaling SR is the default constant scheduler $\gamma =0.5$.

Augmentation	Method	CNN backbone
		R18	R34	R50
Level 0	Original	91.28	92.28	91.57
Level 1	Original	99.10	99.00	98.89
	＋SR (Ours)	99.23	99.27	99.09
Level 2	Original	99.17	99.11	99.00
	＋SR (Ours)	99.27	99.39	99.19
Level 3	Original	99.11	99.06	99.08
	＋SR (Ours)	99.26	99.29	99.20
Level 4	Original	99.12	99.09	99.13
	＋SR (Ours)	$\mathbf{99.40}$	$\mathbf{99.43}$	$\mathbf{99.31}$
Level 5	Original	97.49	97.34	97.66
Level 6	Original	98.50	98.69	98.91