[digiKam-users] face recognition accuracy very poor on large collection

[Travis Kelley]

Hi, I'm running 7.3.0 and have switched to the yolo v3 detection
algorithm.  I have 37k pictures in my gallery that I've accumulated over
the course of many years.  I initially started face tagging with picasa.
Since then I've bounced around to a number of different tools, but have
settled on digikam for image management and pigallery2 for display.

Initially I imported my pictures into digikam which imported some of the
previously tagged face tags, then I tried tagging a few manually, cleared
and rebuilt all training data in the maintenance tab, and finally ran a
recognize faces.  This process took over a week and now I have 32k unknown
faces.  This wouldn't be such a big deal if they were well grouped but when
I look at the person tag for a person and all of the faces that have been
recognized as that person there are more incorrect than correct.  I realize
I've got people like my kids whose faces have changed significantly over
the time I've been taking pictures, but even on adults the recognition
accuracy seems VERY poor with many faces that don't look ANYTHING like the
person they are being recognized as.  Is this expected with the current
face detection algorithms or am I doing something wrong?  Its almost not
worth using the recognize feature on my collection currently since it gets
so many wrong its basically like looking at the unknown group.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.kde.org/pipermail/digikam-users/attachments/20210803/326c8736/attachment.htm>