search

wasidennis / AdaptSegNet

Learning to Adapt Structured Output Space for Semantic Segmentation, CVPR 2018 (spotlight)
847 stars 203 forks source link

Upscaling discriminator output? #82

Closed louis-marchant closed 4 years ago

louis-marchant commented 4 years ago

In the paper, under the discriminator section, you write that "An up-sampling layer is added to the last convolution layer for re-scaling the output to the size of the input."

What is the purpose of this upscaling? How would this affect the loss? I can also see that this upscaling is commented out - did you find out that it was unnecessary?

wasidennis commented 4 years ago

Sorry for the confusion. We did try to upsample the final layer, but the results were not affected. In the final implementation as released here, there is no upsampling layer.