Re: howto improove face recognition rate?

Posted by Maik Qualmann on May 17, 2021; 5:50am
URL: http://digikam.185.s1.nabble.com/digiKam-users-howto-improove-face-recognition-rate-tp4716674p4716682.html

How many faces per person and in total have you assigned so far to train face
recognition?

Maik

Am Montag, 17. Mai 2021, 01:11:58 CEST schrieb KRODER Gerhard:

> Hi,
>
> this is a new 7.2 version, no previous processing/DB data, so no rebuild
> others than starting face detecting/recogintion tool.
>
> Am 16.05.2021 um 22:15 schrieb Maik Qualmann:
> > Have you rebuilt the training database in the maintenance tool (option in
> > the face detection section) if you have already tagged faces with an
> > older digiKam version?
> >
> > Maik
> >
> > Am Sonntag, 16. Mai 2021, 21:16:21 CEST schrieb KRODER Gerhard:
> >> Hi,
> >>
> >> now, some mor data....
> >>
> >> After several runs and tagging face recogintion does not recognize any
> >> more out of already found faces (~25k images, ~6k faces not recognized
> >> ~14k already tagged).  But there are /very much faces very similar/ to
> >> already recoginzed/tagged ones.
> >>
> >> Is there a way for me to icrease recognition rate?
> >>
> >> I've done some "case studies" with my data. Current settings yolo_v3,
> >> with accuracy rate at
> >>
> >>    * 100%  does not suggest any recogniton
> >>    * 95% finds one downscaled/compressed jpg (rate to 10% size=225kB, no
> >>    
> >>      edits)
> >>    
> >>    * 90% one more suggestion but very inaccurate/wronn
> >>    * 85% found 3 more suggestions, one a perfect match, one a fairly
> >>    
> >>      reasonable, one bad
> >>    
> >>    * 80% aims to recoginze/suggest  some more, but with a very low
> >>    
> >>      recoginition rate (17/6000), and low match rate (10/71)
> >>
> >> As i mentioned, having lots of very similar faces assinged already, i
> >> would assume this a poor recognition rate. (6k  pics, ~20% very similar
> >> to already tagged ones, but not found any suggestion.)
> >>
> >>   From the manual, it is suggested to remove tagged "weird" faces
> >>
> >> (carneval or other not very clear representations), to prevent false
> >> suggestions. But this would lead to a  processing and suggestion next
> >> run again, probably false again, wouldn't it?
> >>
> >> Cheers, Gerhard