Fine tuning Object Keypoint Similarity (OKS)

[codingforpleasure]

I have been trying to figure out how to compute the object similarity (OKS), In my custom dataset I have 4 keypoints per image, I'd like during training to improve the Object Keypoint Similarity's score properly, therefore I need to initial the simgas values (cfg.TEST.KEYPOINT_OKS_SIGMAS ) with a list of values (Each simga for each keypoint). I have read thoroughly the Object Keypoint Similarity, BUT unfortunately it's still unclear for me how to find those values:

For each keypoint type i we measured the per-keypoint standard deviation σi with respect to object scale s.
That is we compute σi^2=E[di^2/s^2].

• s we define as the square root of the object segment area

• di are the Euclidean distances between each corresponding ground truth and detected keypoint

it's not clear for me how can I calculate a head of time the Sigmas (Standard deviation) values if the Expected value depends on the Euclidean distances between each corresponding ground truth and detected keypoint, which has not been calculated yet?

Can someone PLEASE give a simple example or explanation so it will be a more clear, Thank you.

[darwinharianto]

From what I know, to make OKS, we need to make our own keypoints on the same image multiple times (make redundant annotations). Pick one as your base, then calculate the euclidean distance and its deviation. This then would be used as sigmas.

[AbhayKarade]

Thanks @codingforpleasure and @darwinharianto, I created my own custom dataset in which keypoints are auto generated, in this case how should I tune the simgas values, do I have to run the auto labeling pipeline multiple times and then calculates the values?