search

dvlab-research / SMR

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

About the regularization term and other assumptions #6

Closed yd-yin closed 3 years ago

yd-yin commented 3 years ago

Hi! Thanks again for your great work!

Your code is clean and easy to understand.

Noticed that there are some regularization terms, and if I'm correct, they are laplacian loss, flat loss and flip loss. It seems that laplacian loss and flat loss are related to the properties of a smooth mesh. What about the flip loss, then? Is it related to symmetry, or is there any assumption about the symmetry in this work?

Thanks a lot!

tau-yihouxiang commented 3 years ago

Actually, at first we didn't use flip loss as a regularization, while we then found that a weakly symmetric assumption can accelerate the convergence.