normal head supervision

[ychgoaround]

Hi,

Thanks for great work! I saw on normal.md there is "For convenience, we did not collect ground-truth normal maps for training. Instead, we derived surface normals from the depth map and camera intrinsics." May I have you tried direct gt normal map supervision for normal head? which design choice will give you better results?

Thanks!