xiyichen / smplify-x-partial

Towards Robust 3D Body Mesh Inference of Partially-observed Humans: semester project at ETH Zurich
Other
59 stars 4 forks source link

Get 3d keypoints from outputs #1

Closed alonelysnake closed 1 year ago

alonelysnake commented 1 year ago

@xiyichen Hi, I'm trying to extract 3d keypoints from the output and using to generate .bvh file, and I encounter several questioins below:

  1. I noticed that 19 points from Expose and 2 points from PIXIE are used as the output, and the results contain 21*3 parameters in the key 'body_pose', but what is the correspondence and hierarchy between the output index and joints?
  2. Is the parameters in 'body_pose' about the parent joint or about the root joint? If about parent joint, is it relative to the prior pose?
  3. There's a key 'global_orient' in the output. I guess it relates to the global coordinate of the root. What does it actually mean and how can I use it to get the global location? Many thanks in advance!
xiyichen commented 1 year ago

Hi @alonelysnake, Thank you for your interest in our work. I believe these questions are more about SMPL-X, and I have found some references for them:

  1. SMPL-X has 21 body joints in axis-angle representation. For the correspondences you can check out here. We initialize the first 19 joints from ExPose (L_Hip to R_Elbow) and the last 2 joints from PIXIE (L_Wrist and R_Wrist).
  2. In SMPL-X the 0th joint (Pelvis) is the root, which I think is eqivalent to the global orientation (check out page 34 here). All other joints are relative to it. We initialize both global orientation and body poses from results of the regression methods. Good point, we should've done a pelvis normalization for the body pose prior, but I believe the effects should be minor since the global orient only changes slightly during the optimization.
  3. The 'global_orient' vector only describes the rotation of the body model in the world coordinates. You can find more details about how to change parameters to new coordinates here.

Hope these answer your questions.