Training with data other than coco about simple-hrnet HOT 3 CLOSED

Hi!
Yes, it could be.

Some thoughts:

• Joints order of COCO and Leeds Sports Pose Extended Training Dataset is different so double check that you have correctly re-ordered Leeds joints to match COCO joint order
• Joints of Leeds Sports Pose Extended Training Dataset seems to be more similar to MPII Human Pose joints. Starting from MPII pre-trained weights may result in higher performance.
• Use a low learning rate and try with different learning rates because it seems to have a high impact on training when starting from pre-trained weights (see issues #26 and #29)

from simple-hrnet.

• Thank you for taking your time. Sorry, missed #26 and #29. Actually decreasing learning rate solved acute drop.
• I'll give MPII a try.
• Do you think missing keypoints (eyes, ears, nose) could be annotated by predicting those images with a pretrained model? Is there a way to recover whether model predicts a point's visibility?

Thanks again

from simple-hrnet.

Do you think missing keypoints (eyes, ears, nose) could be annotated by predicting those images with a pretrained model?

Yes, it could work.

Is there a way to recover whether model predicts a point's visibility?

As far as I know, no because the model doesn't predict that. Actually, the point confidence is usually used to filter the keypoints.

from simple-hrnet.