Keypoint Prediction score

I have successfully built a custom keypoint detection model on a robot. I just have 1 class and 9 keypoints (Head, c1,c2,c3,c4,w1,w2,w3,w4) and its predicting perfectly as well. In situations where the bot is occluded, all the key points are sort of stuffed together within the small bounding box of the detected bot. I thought I will discard the key points with the low scores but I'm not able to figure out the metric used in here. As per the documentation, "pred_keypoints”: a Tensor of shape (N, num_keypoint, 3). Each row in the last dimension is (x, y, score). *_*Confidence scores are larger than 0_.**

My output is something like this:
tensor([[[1.3368e+03, 5.9773e+02, 1.8441e+00],
[1.3268e+03, 6.7090e+02, 5.3277e+00],
[1.3081e+03, 5.9916e+02, 1.6518e+00],
[1.3268e+03, 6.6947e+02, 2.5783e+00],
[1.3066e+03, 5.9773e+02, 1.2810e+00],
[1.2937e+03, 6.8955e+02, 2.0510e+00],
[1.3268e+03, 6.7377e+02, 1.2564e+00],
[1.3268e+03, 6.6947e+02, 1.1330e+00],
[1.3570e+03, 6.0060e+02, 9.5533e-01]]]),
Which when neatly put and assigned with key points are:
head : xposition: tensor(1336.8000) yposition: tensor(597.7300) score: tensor(1.8441)
c1 : xposition: tensor(1326.8000) yposition: tensor(670.9000) score: tensor(5.3277)
c2 : xposition: tensor(1308.1000) yposition: tensor(599.1600) score: tensor(1.6518)
c3 : xposition: tensor(1326.8000) yposition: tensor(669.4700) score: tensor(2.5783)
c4 : xposition: tensor(1306.6000) yposition: tensor(597.7300) score: tensor(1.2810)
w1 : xposition: tensor(1293.7000) yposition: tensor(689.5500) score: tensor(2.0510)
w2 : xposition: tensor(1326.8000) yposition: tensor(673.7700) score: tensor(1.2564)
w3 : xposition: tensor(1326.8000) yposition: tensor(669.4700) score: tensor(1.1330)
w4 : xposition: tensor(1357.) yposition: tensor(600.6000) score: tensor(0.9553)

The score given for each point is what I couldn't understand. Does this simply mean, that the more the score, the more confidence?
I have other Images where the score goes something like 202.32.

Also 'pred_keypoint_heatmaps', parameter seems to be missing in the documentation(https://detectron2.readthedocs.io/en/latest/tutorials/models.html)

Would be grateful if someone could clarify.

Start with the model-output documentation page linked in the issue and trace the documented pred_keypoints output. Verify how the score is defined and whether pred_keypoint_heatmaps is exposed or documented, then update the relevant documentation so both outputs and their interpretation are clear.

Written by the indexing model from the issue text.