scene_generation
scene_generation copied to clipboard
About the loss function of image discriminator in the paper
I read the codes and found there are some differences between the codes and paper on the loss function of the image discriminator. Why the loss function of the D-image in the paper (Page 4)is expressed as:
LD-image=Lreal − Lfake-image − Lfake-layout + Lalt-appearance where Lreal= logDimage(t′, p′) Lfake-image= log(1−Dimage(t′, p)) Lfake-layout= log(1−Dimage(t, p′)) Lalt-appearance= log(1−Dimage(t′′, p′))
In my opinion, according to the codes the loss function should be expressed as:
LD-image=Lreal + Lfake-image + Lfake-layout + Lalt-appearance where Lreal= logDimage(t′, p′) Lfake-image= log(1−Dimage(t′, p)) Lfake-layout= log(1−Dimage(t, p′)) Lalt-appearance= logDimage(t′′, p′)
Maybe I am wrong. Can you tell me if somewhere I misunderstood?