joint confidences vs occlussion + comparison w/ OpenPose

[smidm]

I've got an interesting insight into the keypoints score / confidence values. I have a video of a standing person with a camera rotating around the person. I have plotted the joint confidences for two mmPose models and the OpenPose model. The mmPose HRNET model has confidences bounded in the interval <0;1>, the mmPose MSPN confidences are not upper bounded. Both models confidences behave differently than the OpenPose confidences under occlusion.

• OpenPose confidences drop to zero when occluded, this makes sense for me. Why are mmPose confidences significantly higher?
• Why are the eye confidences so much lower than the ear confidences when fully occluded? I'd expect these to be similar as the OpenPose demonstrates.
• I'd expect that a highly mobile fully occluded joints like wrists should have confidence close to zero. The model is completely making the position up, the joint could be anywhere in the occluded space. Why it doesn't go bellow 0.3/0.4 (HRNET), 0.6 (MSPN)?

Can somebody give me a deeper insight into this?