The reason for represent rotation by quaternion instead of Euler angle. about deep-motion-editing HOT 2 CLOSED

Great question.

Although Euler angles are used in the bvh format, this representation suffers from two main drawbacks:

1. Ambiguity - The same rotation can be represented in various ways.
2. Discontinuity - Values with a large difference (like 0 and 359) may represent close rotation angles.

Hence, this representation is inconvenient for neural networks.

Quaternions representation may suffer from other issues but it’s much less severe. There are papers (like this one) that discuss these issues and try to propose alternative representations for rotations in neural networks. Empirically, to us, quaternions work well.

from deep-motion-editing.

Got it! Thanks for your explanation!

from deep-motion-editing.