Meaning of keypoints_scores (Keypoint R-CNN)

Hi all,
I am training a Keypoint R-CNN model and, at inference, I get, among others, keypoints_scores.
I had assumed those were the “confidence scores” for each keypoint, same as scores for bboxes.
I was surprised to see, though, that keypoints_scores can equal any value > 0 (e.g. I am getting 10 or 20) and are not constrained in the range 0-1, as bboxes scores are instead.
I could of course normalize those via a sigmoid, but I just want to make sure I understand what those are before doing anything stupid.
Given I don’t see any relevant documentation on keypoints_scores, can anyone please shed some light?
Thanks a lot

1 Like