The mediapipe code identifies 21 hand points, while the OpenXR specification has 25. Currently the "extra" OpenXR points are left invalid, but it would be better if the mediapipe code could be updated to identify all 25 points. This could be substantial as it may require changing the hard coded tensorflow-lite models.