How do you process the "joint_3d_image" in the datasets

[hs1ang-hsu]

In your datasets, the "joint_3d_camera" means the 3D pose in camera coordinates. The "joint_3d_image" should be the 3D pose in a 2D projective plane, which can be obtained by applying the intrinsic matrix to the vectors in "joint_3d_camera." However, I can not figure out how you get the values in the z-axis. Would you please clarify the process? Thank you!

[Embracing]

Thanks for your interest in our work. Please refer to this code for how to get the z-value.

[hs1ang-hsu]

Thanks for your reply. The code is very helpful!
It seems that your dataset assumes known intrinsic parameters and known depth of the root joint, which simplifies the 2d-to-3d lifting task. Other works use normalized 2d keypoints to estimate 3d joints in camera coordinates. In your dataset, you already applied the intrinsic parameters on 2d keypoints. And the depth of ground truth 3d joints is multiplied by f_x / z_0 + 1/2000, where f_x is the focal length and z_0 is the z-value of the root joint.
Your method introduced in GFPose is amazing. Thus I am very interested to know the performance on the widely used settings of 2d-to-3d lifting task. Have you tried your method in this case before? If so, would you mind to share the result here? Thank you very much!