Add way to "filter out" low confidence keypoints

[Dref360]

Search before asking

• I have searched the Supervision issues and found no similar feature requests.

Description

In pose estimation, low confidence keypoints should be set to 0.
It would be great to include this in supervision instead of asking the user to do it.

Proposal:

kpts = KeyPoints(xy=...,
                 confidence=...,)

kpts = kpts.with_threshold(threshold=0.5)

All keypoints with confidence lower than 0.5 would be set to 0.

Use case

Useful when visualizing skeletons with low confidence.

Additional

No response

Are you willing to submit a PR?

• Yes I'd like to help by submitting a PR!

[miteshashar]

@Dref360 I am trying to better understand the utility of this.

Setting the key points below the threshold to 0(if that is what you mean) could in some cases result in a low %age of key points remaining in place – the ones satisfying the threshold.
The key points determine the shape of the pose. And hence, this would effectively result in just a less sparse pose structure. Which is why I am unable to grasp what you are trying to visualize.

To visually understand what is the impact on the skeletons when I mute key points below a certain threshold, I ran a small experiment to visualize it.

Below is a sample of these illustrations, there are more on the repo.

3.webm.mp4

What would you do with the KeyPoints object having key points below the confidence threshold muted? It would be great if you can point to some plausible use cases you are trying to achieve to clarify whether I am misunderstanding what your requirement is.

Also, now with slightly better initial understanding of Supervision's code, I feel the method you have proposed is a very thin abstraction and can be achieved simply with one line of code, since kpts.xy and kpts.confidence are just NumPy arrays:

THRESHOLD = 0.5
REPLACEMENT = 0
kpts.xy[kpts.confidence < THRESHOLD] = REPLACEMENT