FengZhenhua/Wing-Loss

Question about PDB

mrgloom opened this issue · 3 comments

@mrgloom Hi, in general, the 1st eigenvector controls pose variations. But, I visualised the results by changing the weights of the first 3 eigenvectors for different datasets. I found that, for AFLW, the 2nd one controls pose variations. This is why I used the 2nd for AFLW and 1st for the others.

What does it mean 'controls pose variations' in terms of pitch, yaw, roll?

@mrgloom Hi, here the term 'pose' is a general term. But, in our method, we focus on the yaw rotation for the term 'pose'. You can also do data balancing for any other `poses' or a mixture of different 'poses'.