details on 'facepose' outputs

Dear all,
I would like to create rotation matrix with facepose outputs (i.e. pitch, yaw, roll) to apply rigid inverse affine translation so that I can apply rigid motion correction. However, because the landmarks are in 2D, I do not know how to do this calculation.
I thought of getting some clues from the developers on how they calculated the facepose values but I could not find any information. Could you please provide some details?
Thanks!

Start by reviewing the facepose outputs—pitch, yaw, roll—and the 2D landmark limitation described in the issue. Document how these values are calculated and whether they are sufficient for constructing a rotation matrix and rigid inverse affine transformation; the issue is complete when those details and relevant constraints are clearly explained.

Written by the indexing model from the issue text.