Confusion about the canonicalization.

[hoyeYang]

Great work! But I'm confused about these lines

FlowMDM/data_loaders/amass/transforms/rots2rfeats/globvelandy.py

Lines 69 to 73 in e81c67f

	rot2d[..., :2] = 0
	if self.offset:
	# add a bit more rotation
	rot2d[..., 2] += torch.pi/2

Assuming the original rot2d is $(x\theta, y\theta, z\theta)$, directly transforming the rot2d represented by axis-angle into $(0, 0, z\theta+\frac{\pi}{2})$ seems confusing.

If the canonicalization is removed, will the experimental results be significantly affected?