How to improve the quality of keypoints 2D and 3D?

[chandelarvind]

I have tested on rtsp live feed to get 2D and 3D poses. but i see so much variation on keypoints extracted due to which even 2D visualization is not looking good. How can we make it better, is there any better model file available. and what is role of Focal length and principal points in computation because i see they are commented in pix2cam function.