Issue about the pixel-wise discriminator

wasidennis / AdaptSegNet

Learning to Adapt Structured Output Space for Semantic Segmentation, CVPR 2018 (spotlight)

847 stars 203 forks source link

Issue about the pixel-wise discriminator #64

Closed HanqingXu closed 4 years ago

HanqingXu commented 5 years ago

Hi, I check the code of the discriminator, and find that the output is not the same size as the input(you comment the upsampling code). So I wonder if the resolution of the output would make a differnence in your practice, and its difference compared with just binary domain classifier (that is , just single-value output). Thanks in advance.

hfslyc commented 5 years ago

Hi,

Without the upsampling layer, the output of discriminator is still a 2D map and is not a single value. We found that upsampling the discriminator output does not affect the performance much but introduce slight computation overhead.