Human3.6M 2D to 3D coordinates

[vikingsheadquarter]

Hello,

I have downloaded and preprocessed the Human3.6M dataset for a Uni project and have the annotations in COCO format.
I am new to 3D pose estimation, so I am trying to figure out the conversion of 2D image coordinates to the 3D coordinates, as the json file provides 'keypoints' and 'keypoints_3d'.
From my understanding, the 3D coordinates are given relative to the bounding box and to the root bone.
However, I have not succeeded to understand how 2D predictions need to be processed for 3D lifting.

Any help is highly appreciated!