Page 1 of 1

Extraction question about object on the face & weird too-wide jaw line alignment

Posted: Sun Feb 04, 2024 9:10 pm
by sonicdriller07

Hi, I'm extracting training set data from a movie, and I used Bisenet-Fp masker during extraction.

  1. Do you include picture(s) of face that some part of faces are hidden by objects (like book, leaves, tree, or wall...) to training data?
    I usually exclude if face is covered by objects more than 50%, but not sure how others do.

  2. During manual fixing stage, sometimes I see the weird alignment where the one side of jaw too wide than actual too wide, but eyes, nose, and mouse detection & alignment are fine.
    I checked other post of similar topic (viewtopic.php?t=1677), so the one of the answer mentioned it depends on mask; landmarks based mask it matters so I need to fix jaw line, but if it is not landmark based mask then it doesn't matter so ignore the jaw line.
    As I mentioned above, I used Bisenet-Fp masker during extraction. Is Bisenet-Fp masker landmarks based mask, so I need to fix them all? I've been manually fixing those weird jaw lines but it takes too much effort & time...


Re: Extraction question about object on the face & weird too-wide jaw line alignment

Posted: Mon Feb 05, 2024 12:34 pm
by torzdf
sonicdriller07 wrote: Sun Feb 04, 2024 9:10 pm

Do you include picture(s) of face that some part of faces are hidden by objects (like book, leaves, tree, or wall...) to training data?

If they are masked correctly, I do include them, yes.

sonicdriller07 wrote: Sun Feb 04, 2024 9:10 pm

As I mentioned above, I used Bisenet-Fp masker during extraction. Is Bisenet-Fp masker landmarks based mask, so I need to fix them all? I've been manually fixing those weird jaw lines but it takes too much effort & time...

I call this 'fly-away jaw'. Don't worry about them. Jaw landmarks aren't used for anything if you are using Bisenet-FP