Question about keypoint supervision and performance

[hebing-sjtu]

Hi, awesome work on this project! 👏

I’m curious about the performance difference when using the model originally trained to predict m keypoints (e.g., 133) versus one trained for n keypoints (e.g., 17).

Specifically:
• Have you observed any differences in the accuracy or quality of the n keypoints when using the m-keypoint model?
• Does the additional supervision from the extra keypoints help or hurt the prediction of the backbone n keypoints?

Thanks in advance for any insights!