The view consistency of the monocular video reconstruction results using VHAP + Gaussian Avatars is very poor

[Haoawei]

I used VHAP to process the monocular video in this link(https://syncandshare.lrz.de/getlink/fiJE46wKrG6oTVZ16CUmMr/VHAP), then input the results into Gaussian Avatars for training. The results show that the front-facing reconstruction quality is acceptable, but the side-facing results are very poor when rotating. Strangely, however, the side-facing results are OK when only the head is rotated. Could you please tell me the specific reasons for this? And what solutions are there to fix it?

[ShenhanQian]

It is advised to set sh_degree to 0 for monocular input. The default value is 3.

GaussianAvatars/arguments/__init__.py

Lines 47 to 49 in 30cc4bd

	class ModelParams(ParamGroup):
	def __init__(self, parser, sentinel=False):
	self.sh_degree = 3

[Haoawei]

It is advised to set sh_degree to 0 for monocular input. The default value is 3.

GaussianAvatars/arguments/init.py

Lines 47 to 49 in 30cc4bd

class ModelParams(ParamGroup):
def init(self, parser, sentinel=False):
self.sh_degree = 3

That's good. It works but WHY? In my view, the expressive ability should be better with sh = 3. Will setting it to 0 compromise the reconstruction quality?