Why is SMPL pose condition zeroed out before 20 epochs and every 20 epochs after that ?

[vhvkhoa]

Hello, thank you very much for your work and published code.

I noticed at the code snippet below:

vid2avatar/code/lib/model/v2a.py

Lines 106 to 108 in a1ab86a

	if self.training:
	if input['current_epoch'] < 20 or input['current_epoch'] % 20 == 0:
	cond = {'smpl': smpl_pose[:, 3:] * 0.}

SMPL condition for SDF network is zeroed out before 20 epochs and every 20 epochs afterwards. Can I know why is it so ?

[MoyGcc]

Hi, thank you for your interest! It's actually a simple trick to avoid overfitting of the canonical SDF network to certain human poses. This helps to keep a more consistent canonical human shape. For example, more back details can be preserved for 3D human reconstruction in frames where the human is facing the camera frontally.