Selmaan

I had the same problem trying to run the 'batch processing' script with only 'quality metrics' module uncommented, and downgrading fixed it for me as well

Has there been any update to this? Can you describe what challenges there are in implementing this? I'm not particularly experienced, but this is critical for me and I'd be...

Ok thanks please let me know!

Thanks for this! I verified that it works, in the annotator (but not downstream training). I'm not sure I totally understand the concern around training. As I understand, 'non visible'...

I think I understand your points now, except that I don't understand why making essentially random predictions for non-visible key points will impact the euclidean error? If the key point...

I've verified that training works for non-visible points, thanks for making the change! I disabled graph prediction and observed wacky euclidean errors as you said. Something that confused me at...

So I was going to implement this, but then realized it could be a bit tricky. There are three places changes could be implemented: (1) Logger callback on_epoch_end; (2) base_model.evaluate;...

Ok, I just wasn’t sure because if we are modifying utils.keypoints.keypoint_errors function, I worried it might be used at other places in the code. So it sounds like the thing...

Just made a pull request for this issue, plus a few minor things that came up in implementing it

This push is to fix the issues discussed in open issue #27