I'm trying to filter out bad face images (with artifacts) generated by StyleGAN.
Is there a good image quality measurement mechanism in python for doing this?
I've tried using BRISQUE but it is not showing good results (there are a lot of good quality images with higher scores then for bad quality images).

EDIT: These artifacts could be any kind of distortions, like eyes looking in a different directions, asymmetries (e.g. could be bad looking ear, assymetric eyes, ...), strange skin coloring, image messed up with the background like below image.distoretd image

More images could be found here.

  • $\begingroup$ it would probably help if you could define these "artifacts" somehow. $\endgroup$ – Marcus Müller Sep 26 '19 at 14:03
  • $\begingroup$ @MarcusMüller I've edited my question with more details. But unfortunately I cannot define artifacts in a proper manner. I've tried also face landmarks detection, but still it succeeds on such distorted images. $\endgroup$ – hemel Sep 26 '19 at 14:18
  • $\begingroup$ Um, so your description is "something that seems perfectly plausible to a complex neural network trained on basically all statistics of an image class but isn't plausible to the human observer"? because bad news: you will have to write something that surpasses the GAN in being able to build images of faces... $\endgroup$ – Marcus Müller Sep 26 '19 at 14:26
  • $\begingroup$ (also, honestly, squinting a bit: that is the picture of a man with crossed eyes and a large scar on his left cheek in half-profile, not so "artifacty", to me) $\endgroup$ – Marcus Müller Sep 26 '19 at 14:28

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.