Question about reference paper

[John1231983]

Thanks for your implementation. Do you implement the scheme of paper "Semantic Segmentation using Adversarial Networks" https://arxiv.org/pdf/1611.08408.pdf? Could you tell me how much gain did you achieve when using FCN and FCN+GAN?
Thanks in advance

[rogertrullo]

Hi @John1231983 , yes it is the same idea of that paper. A difference is that in the discriminator I only use the output of the generator. In the mentioned paper they concatenate the output of the generator with the input image. My first experiments didn't show improvements, but I didn't
try to hard finding the right params, and neither I tried using the concatenation as in the original paper. It should be easy to add to the current code though.

[John1231983]

Me too. I did not see any improvement when ran the GAN+FCN and FCN only.

[John1231983]

Hi @rogertrullo: Do you have any update results? I am wondering about the benefit of GAN for segmentation