The inverse kinematic doesn't reflect true orientation of a motion.

[yujuisu]

The orientation of each bone are computed downstream from a chain:

for chain in self._kinematic_tree:
    R = root_quat
    for j in range(len(chain) - 1):
        # (batch, 3)
        u = self._raw_offset_np[chain[j+1]][np.newaxis,...].repeat(len(joints), axis=0)
        # print(u.shape)
        # (batch, 3)
        v = joints[:, chain[j+1]] - joints[:, chain[j]]
        v = v / np.sqrt((v**2).sum(axis=-1))[:, np.newaxis]
        # print(u.shape, v.shape)
        rot_u_v = qbetween_np(u, v)

        R_loc = qmul_np(qinv_np(R), rot_u_v)

        quat_params[:,chain[j + 1], :] = R_loc
        R = qmul_np(R, R_loc)

But consider these:

1. Lift the arm by rotating the shoulder. You should update the orientation of the arm. But instead you update the relative rotation between forearm and arm.
2. It can’t detect twist around the forearm axis. Suppose the wrist rotates (e.g. turning a screwdriver), but the elbow and wrist positions stay fixed.

[yujuisu]

You process data by smpl, the model uses axis-angle already. Could you explain why you ignored this?

[YYYYYHC]

Hi,

Do you think my issue #178 is relevant here?

[yujuisu]

If you understand what I wrote, you'd switch to using data from AMASS directly.

[YYYYYHC]

I see. I've switched to another representation (https://zju3dv.github.io/MotionStreamer/) that doesn't rely on IK. Thanks for your reply.