How to map 3d keypoints (openpose) to smpl inputs (betas, trans)?

[szang18]

Hi, thank you for the great work!

I have a question regarding the SMPLer-X image generation. In your paper, you show visualizations of SMPLer-X results, which seem to be generated from 3D keypoints. I’m trying to reproduce this, but I couldn't find the exact code or the pipeline for generating those SMPLer-X images from 3D keypoints since SMPLer-X requires pose parameters (i.e., relative joint rotations)

Could you kindly point me to:

• The relevant code/script for this process

• The steps required to convert 3D keypoints into the SMPLer-X rendered outputs

Any guidance would be greatly appreciated. Thanks in advance!

[SignLLM]

@szang18 For the SMPL part, I used a diffusion model to render the generated pose videos into real videos and then conducted the evaluation. It is not the main content; it is merely to demonstrate our temporary applicability. To be honest, this part is a bit redundant. Different researchers have the freedom to use whatever method they prefer for pose extraction. The current limitation of SMPL is not because the pose extraction method is inferior to that of the skeleton pose model; it is simply because this field needs to develop and computing power needs to be improved.