Identity PCA modes are not orthogonal ?

[spaceapple]

Hi,
Thanks for the awesome work, the data and code are super useful.

Something is bothering me though. It's my understanding that the identifyXXX.obj files store the identity PCA modes.
So, I'd expect those to be orthogonal to eachother.
In code

import face_model_io
face_model = face_model_io.load_face_model('../FaceXModel')
mode_dotproduct = np.sum( face_model._identity_shape_modes[0] * face_model._identity_shape_modes[1])

I'd expect mode_dotproduct to be zero (or very close to it), but it's not the case.
I get

>>> mode_dotproduct
0.0005141700283957107

The same goes for all combinations of shape modes. So they're not PCA eigenvectors ?

[SunMoonDev]

maybe the obj is not complement equal to the original pca result. maybe there are some postprocess apply to the original pca result