Skip to content

Why is camera coordinate data passed to SMPLLoss during computing MotionLoss? #51

@linjiangya

Description

@linjiangya

Dear authors,

In SmoothOptimizer and RootOptimizer, I think the joints3d/joints3d_op in SMPLLoss/RootLoss is in the world coordinate system. And the re-projection is performed to get the proj_2d joints from (world -> pixel coordinates).

However, it seems like in MotionLoss, the cam_pred_data is passed into RootLoss:

loss, stats_dict = super().forward(
observed_data, cam_pred_data, nsteps, valid_mask=valid_mask
)

Whilst I think from the logic in rollout_latent_motion(), it looks like cam_pred_data is actually still in the world coordinate system, and is transformed back from the local coordinate system used in the rollout procedure. I am therefore confused about the name of the variables: Why is this data named cam_pred_data with a comment above saying "must return trans and root orient in camera frame...", while the function name is apply_world2prior? And it doesn't look like camera coordinate system data to me.

I just would like to confirm whether this is a naming issue or I misunderstood it. Thank you!

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions