search

yinguobing / facial-landmark-detection-hrnet

A TensorFlow implementation of HRNet for facial landmark detection.
GNU General Public License v3.0
156 stars 40 forks source link

NME with original image dimensions #15

Closed Sheepy0 closed 3 years ago

Sheepy0 commented 3 years ago

Hello, Thank you @yinguobing for your work and blog, I am really enjoying them.

In the evaluation, before we compute NME, we to transform the marks back to the original image dimensions. Do you know why the transform of the marks back to the original image dimensions is necessary for the NME ?

Thank you

yinguobing commented 3 years ago

Actually I didn't pay much attention to the NME computation. In the code the predictions are transformed back because the labels are of image dimensions. If the predictions won't be transformed, then we will have to transform the labels into the cropped image dimension.

Sheepy0 commented 3 years ago

Great, this makes sense ! Also, at the moment, I think the NME computation is only being performed on a single image per batch, rather than on the entire batch.

Thanks for your quick response, have a good day.