jason718 / game-feature-learning Goto Github PK

Hello jason, thanks for your hard work.
It seems that the synthesized data is assigned to 0 in the code and real data is set to 1, which is different from the paper.
self.loss_D_syn = self.criterionGAN(pred_syn, False)
self.loss_D_real = self.criterionGAN(pred_real, True)
I'm new to the field of domain adaption and GAN. I don't know if this will affect the performance.
Looking forward your response. Thank you.

First of all, thanks so much for your amazing work!

I have a question on how you detach synthetic features. Detaching the synthetic feature during the update of D makes sense to me, but In this line, you detach the synthetic feature during updating the B and H. If you do so, then network B will not be able to learn domain-invariant features because the weights won't be able to be backproped to B because of detaching.

Just curious about your explanation.

Thanks so much for your help in advance.

Thanks for sharing the code.

I couldn't find any info regarding where to get the ground-truth edges. In SceneNet, there is a script for generating normals from depth data. But how are edge images generated?