There is an issue with the skeletal data annotator. More specifically inside of the python file OgnGetSkeletonData.py within the compute_2d_translations function the logic for determining if a skeleton is inView of the camera appears incorrect,
# Check if the current skeleton is in view of the viewport
skel_data[“in_view”] = not np.any(
np.any(joint_pos2d[:, 0] < 0)
and np.any(joint_pos2d[:, 0] > view_params[“width”])
and np.any(joint_pos2d[:, 1] < 0)
and np.any(joint_pos2d[:, 1] > view_params[“height”])
)
I believe these and statements should actually be or statements. In addition this is classifying if the entire skeleton is in view based on if even one joint is out of view, the docs state that there should be a bool returned for each joint that tells you if it is in view of the camera. In addition there should be a check added to see if points are behind the camera, right now when they are projected to 2d from 3d the logic will classify points behind the camera but within the view frame when the dimension is squashed as inview, when in reality they are not visible to the camera.