search

hongsukchoi / Pose2Mesh_RELEASE

Official Pytorch implementation of "Pose2Mesh: Graph Convolutional Network for 3D Human Pose and Mesh Recovery from a 2D Human Pose", ECCV 2020
MIT License
678 stars 69 forks source link

Keypoints Conversion - SMPL to COCO #70

Open Dipankar1997161 opened 1 year ago

Dipankar1997161 commented 1 year ago

Hello @hongsukchoi,

Hope you are doing well.

I generated the Smpl and successfully run the 3d reconstruction on those points.

My final task is - extracting the smpl key points (which are the original smpl joints) and then performing 2d and 3d pose estimation on the data.

Q1. The smpl key points are the "Joints" or the "Poses" parameter in Smpl? Q2. How do I extract the smpl key points and convert them to coco format?

Can you provide some insights on this? Any repo or Pose2Mesh can provide any such thing?

I am okay with both methods but haven't found many reliable sources for the conversion of SMPL to COCO.

I will wait for your response.

hongsukchoi commented 1 year ago

Hi @Dipankar1997161, sorry for the late reply.

Q1. The smpl 'pose' parameters are joint angles (not joint position) in axis-angle representation. The smpl key points you are referring to are probably the "Joints" (joint position). Q2. You could estimate SMPL pose paramters from the smpl keypoints and reconstruct the 3D mesh, and multiply the coco joint matrix to the mesh vertices to get the coco joints. coco joint matrix: https://github.com/hongsukchoi/Pose2Mesh_RELEASE/blob/master/data/COCO/J_regressor_coco.npy This joint matrix is approximately performing forward kinematics.