Animation: sampler.interpolation values

[pjcozzi]

AFAIK the only valid interpolation in glTF 1.0 is LINEAR, e.g.,

"sampler": {
  "input": "TIME",
  "interpolation": "LINEAR",
  "output": "rotation"
}

For rotations, both @tparisi and my implementations use slerp, which is "spherical linear" so don't we think LINEAR is misleading?

Perhaps we

• Introduce SPHERICAL_LINEAR for rotations, or
• Drop interpolation from the 1.0 spec and say that all values are LINEAR except a node.rotation target, which would be SPHERICAL_LINEAR.

Thinking forward to material targets that may have rotation matrices as animated parameters, we may want SPHERICAL_LINEAR here (although it implies conversions to/from quaternion) since linear interpolation of rotation matrices does not maintain an orthonormal matrix except at the key frames. Having an explicit SPHERICAL_LINEAR would also avoid the case of trying to apply slerp to a scale matrix just because it is 3x3.

[fabrobinet]

That's a possibility, but I would say only if we agree here: #144 that we support quaternion explicitly.

[pjcozzi]

only if we agree here: #144 that we support quaternion explicitly.

OK. I suppose this adds some pressure to supporting quaternions then.

[pjcozzi]

@fabrobinet do you recall what interpolation the converter expects for rotations?

[pjcozzi]

@tparisi to close this issue, I suggest we add the following wording to the spec:

When an animation targets a node's rotation and the animation's interpolation is LINEAR, spherical linear interpolation (slerp) should be used to interpolate quaternions.

[pjcozzi]

This is in the spec now.