questions about perceptual loss weight

vinthony / ghost-free-shadow-removal

[AAAI 2020] Towards Ghost-free Shadow Removal via Dual Hierarchical Aggregation Network and Shadow Matting GAN

https://arxiv.org/abs/1911.08718

297 stars 59 forks source link

questions about perceptual loss weight #14

Closed howardgriffin closed 4 years ago

howardgriffin commented 4 years ago

Hi, I am confused of the different weight of each vgg layer's perceptual loss , and the GAN loss weight seems to be very low compared with other loss. Could you make some explanation?

vinthony commented 4 years ago

Hi, thanks for your attention.

The weights of the perceptual loss are taken from the previous work on reflection removal.

GAN plays as a regularization for our task and aims to fit the overall distribution. Thus, if we use a larger GAN loss, the output will not be produced as expected compared to the ground truth. and, we still want our network to mainly focus on pixel-wise restoration. The weight of GAN loss follows Pix2Pix and more details are discussed in that paper.