Question about global pose

[huzijun1996]

I synthesized the global rotation data as you implied in preprocess.py.
for i, l in tqdm(list(enumerate(length))):
if l <= 12: b += l; print('\tdiscard one sequence with length', l); continue
p = art.math.axis_angle_to_rotation_matrix(pose[b:b + l]).view(-1, 24, 3, 3)
grot, joint, vert = body_model.forward_kinematics(p, shape[i], tran[b:b + l], calc_mesh=True)
out_grot.append(grot[:])
torch.save(out_grot, os.path.join(paths.amass_dir, 'grot.pt'))
I directly used inverse_kinematics_R on the file to get the local rotation data as shown in Figure 1 with no problem.

But if I select only 15 of those nodes and use def _reduced_glb_6d_to_full_local_mat(self, root_rotation, glb_reduced_pose): to generate 24 node data
reduced = [1, 2, 3, 4, 5, 6, 9, 12, 13, 14, 15, 16, 17, 18, 19]
ignored = [0, 7, 8, 10, 11, 20, 21, 22, 23]
That birth result is shown in Figure 2, where the waist joints are not harmonized with the body.

May I ask what you did with the global rotation data in addition to the synthesis?

09c5e3dfb76177483f5257f6d1897e6f.mp4

7bf3af2870ca2cafad7c539df495b037.mp4

[Xinyu-Yi]

During training, we use the rotation of the 15 joints expressed in the root frame. So you need to left-multiply the inverse of root rotation to the world-frame global rotation and then select the 15 joints.

[Xinyu-Yi]

i.e. the training gt is something like grot[:, :1].transpose(2, 3).matmul(grot)[:, reduced_joint_indices]

[huzijun1996]

Thanks. In the joint position processing in joint.pt, I'm solving for the relative positions of leaf and full joints: leaf_joints_position - root_position and full_joints_position - root_position. Is there any preprocessing needed other than that?

[David202307]

@huzijun1996 能加个微信吗，请教一下13450294131