It seems that the Attention Network can pay attention to foreground object essentially, without lable information about attention region.

[jian3xiao]

Thanks for your work and i have read your paper. I have a question about the attention network.

In unsupervised setup, the losses are adversarial loss and cycle-consistency loss during early of training without label information abut foreground object (for exampel, binary labels of segmentation as attention map). How the attention network focus on foreground object ? And the network architecture of Generators and Attention Networks are almost same except the final activation function. Why the final activation function is sigmoid with output channel is 1, output of the network is considered to be attention map? While another ouptput is considered to be reconstruction of input. I don‘t know how that works.

Moreover, Figure 7 in the paper shows Attention Network can focus on foreground object in early of training. It seems that the Attention Network ( or CNN ) can pay attention to foreground object rather than background essentially, even without lable information about attention region.

[jacobw]

To my understanding and if i understand your question correctly, to fool the source discriminator as in conventional GAN, the attention work is required to attend to the most domain discriminative regions other than other sub-patches. And the foreground is believed to be the most domain descriptive. Regarding to output channel 1, the attention map is a single channel mask with continious values between 0 and 1.