search

dvlab-research / SMR

Self-Supervised 3D Mesh Reconstruction from Single Images (CVPR2021)
93 stars 8 forks source link

Questions about GAN architecture and 2D-only supervision #4

Closed yd-yin closed 3 years ago

yd-yin commented 3 years ago

Hi! Thanks a lot for your great work, and thanks for releasing the code! I got some questions.

In the paper, I did not find the GAN architecture, while there is GAN loss in the code. Is GAN loss also part of this work?

\ Another question is about the implementation. Can I re-implement the 2D-only supervision by this code?

I get an error if I set lambda_ic to 0, since Xir is rendered all the time, https://github.com/dvlab-research/SMR/blob/master/train.py#L547 and Ai and Aire are used in losses, e.g., https://github.com/dvlab-research/SMR/blob/master/train.py#L577

Thanks a lot!

tau-yihouxiang commented 3 years ago

Yes, we fould that GAN loss was pretty useful during training. For 2D-only supervision, I will check it soon when lambda_ic = 0.