Question about batch["target_motion"]/["source_motion"] composition and joint order in MotionFix

[neoory]

Hi! Sorry to bother you again!

I have a question regarding the motion representation:
In configs/model/basic_clip.yaml, I see that the dataset uses the following input_feats:

input_feats:
- "body_transl_delta_pelv"
- "body_orient_xy"
- "z_orient_delta"
- "body_pose"
- "body_joints_local_wo_z_rot"

However, during training, I noticed that:

batch["target_motion"] != torch.cat([
    batch["body_transl_delta_pelv_target"],
    batch["body_orient_xy_target"],
    batch["z_orient_delta_target"],
    batch["body_pose_target"],
    batch["body_joints_local_wo_z_rot_target"]
], dim=-1)

(and similarly for batch["source_motion"])

Does this mean that batch["target_motion"]/["source_motion"] is not simply a concatenation of input_feats? If so, could you clarify how motion_target is constructed from the features?

---

By the way, I have a question regarding the motion features. Specifically, I would like to know the joint order for:
"body_pose" (21 joints*6)
"body_joints_local_wo_z_rot"(22 joints*3)
I assume these correspond to the SMPL-X model, but could you please confirm the exact joint ordering for these features?

Thank you very much for your help!

[atnikos]

Hey @neoory ,

Thanks for your interest :)
The representation is a concatenation of these features. Their order should be the same for both source and target motion.
This is done in this part of the code.

Regarding your second question, the body joints have the pelvis excluded (that is why you are seeing 1 less). The pelvis is encoded separately via these:

- "body_transl_delta_pelv"
- "body_orient_xy"
- "z_orient_delta"

The local joints are the joint positions.

[neoory]

Thanks for your reply!

For Q1,
I checked the batch when training, and the structure of batch as follows:

It can be observed that the beginning of batch["target"]/batch["source"] is not the same as each self.input_feats, i.e., there are no data "0.0013" or "1.3433e-03" at the beginning of batch[self.input_feats].

For Q2,
So, my understanding is:

• The joint sequences of "body_pose" with SMPLX is [1,2,3, ...,21]
• The joint sequences of "body_joints_local_wo_z_rot" with SMPLX is [0,1,2,3, ...,21]

which "body_pose" and "body_joints_local_wo_z_rot"'s joint numbers are the same as the SMPLX joint numbers.
Is it right?