Cases where 3D-pose-baseline falls short The model is trained on a dataset with relatively limited variation:
Recorded in a studio Cameras at a few specific angles A few different models For some scenes, the 3D skeleton seems to be leaning forward where it shouldn’t. My 2D predictions are from OpenPose, which don’t have the same skeletal structure. In particular, these points are missing:
Hip (there are left and right hip points, though) Head (as in somewhere at the top of the head) Spine (somewhere chest-level) So I have to fill these in myself and play around a bit with what would be a good heuristic.