Question about revert_axis_swap about articulated-animation HOT 4 OPEN

When you compute pca orientation of the axis could be arbitrary. For example x axis in canonical space may point left or right. If the orientation is different in the source and driving, corresponding region will be flipped.

from articulated-animation.

Thanks for your reply. Does this only prevent the horizontal flip? It seems only the first basis is considered.

For preventing vertical flip, do we also need to do value = value * torch.sign(value[:, :, 1:2, 1:2])?

from articulated-animation.

I did not observe problems with other flip. So I did not include this, this probably depends on actual svd implementation.

from articulated-animation.

I have encountered some strange flips even with value = value * torch.sign(value[:, :, 0:1, 0:1]).

Yes, it should relate to the svd implementation. Just found torch.svd (PyTorch 1.8.1, not sure whether other versions have this problem) does not guarantee that the determinant of u always has the same sign.

The following workaround together with value = value * torch.sign(value[:, :, 0:1, 0:1]) solved my problem:

# flip the first basis when det < 0
det = torch.linalg.det(u).unsqueeze(-1).unsqueeze(-1)
u = u * torch.sign(torch.cat([det, det.abs()], dim=-1))

from articulated-animation.