Request for Clarification on pose to pose_rescaled Conversion Logic

[elliottzheng]

Hi, thanks for your great work! I'm really impressed with this project.

I'm currently working on a related task and have a question regarding the keypoint processing pipeline. I've successfully used smooth_kp.py to get pose_rescaled_smooth from pose_rescaled and then converted it to local coordinates using the global_to_local function.

However, I'm stuck on the initial step of generating pose_rescaled from the raw pose data. I've searched the repository but haven't been able to find the specific logic or code for this rescaling step.

Could you please provide some guidance on how pose is converted to pose_rescaled? Is there a specific function or script I might have missed? Any insights into the logic or a snippet of the relevant code would be greatly appreciated.

Thanks a lot for your help!

[xyjin01]

+1
Additionally, I'm unsure about how to generate the 133x3 'pose' from dwpose. I suspect it involves removing the neck from the 18 body keypoints, then reordering them to match the COCO format, followed by concatenating with face and hands. However, the range of 'pose' seems inconsistent compared to dwpose; the former might span from 0 to H/W. Is this obtained by multiplying dwpose outputs by the height and width?

Such excellent work would be even more remarkable if the data preparation section were clearer.