search

wasidennis / AdaptSegNet

Learning to Adapt Structured Output Space for Semantic Segmentation, CVPR 2018 (spotlight)
845 stars 205 forks source link

when I removed the discriminator I found the results GTA5->Cityscapes degraded by 1% only #95

Open MoHassoubah opened 3 years ago

MoHassoubah commented 3 years ago

Hi, First, thanks for your paper and your contribution and making you work available to everyone to benefit from it.

I trained the network (GTA5->Cityscapes vanilla gan) without backpropagating the adversial loss in the generator, this way image and I got these validation results without the discriminator (max mIOU=41.79) orignal network performance without discriminator (with low disc loss)

and these are the validation results with the discriminator effect (max mIOU=42.45) orignal network performance -with- discriminator

the number on the left are the number of iteration at which the weights were saved and the numbers on the right are the mIOU at each weights snapshot

I find the discriminator effect is little is that normal? or it should improve the results more than that?

MoHassoubah commented 3 years ago

Hello @wasidennis would you please answer my question