Table 6:

To evaluate the effectiveness of the patch-wise attention, we compare the proposed model with the variant (uniform) that always assigns equal attention to all patches. To investigate the importance of the localization information in the saliency maps, we trained another variant (random) that randomly selects patches from the input image. We use GMIC-ResNet-18 model with top 3% pooling as the base model. The performance of the local module (ŷ_local) is reported.

Attention	ROI patches	AUC(M)	AUC(B)
uniform	retrieve_roi	0.874 ± 0.008	0.776 ± 0.007
gated	random	0.629 ± 0.042	0.658 ± 0.011
gated	retrieve_roi	0.898 ± 0.01	0.78 ± 0.008