[digiKam-users] Questions about face recognition and workflow

[Jens Benecke]

Hello,

I am currently cleaning out my photos database of ~150'000 images, taken 
in the last 20 years roughly.

Until 2015, this database was maintained using iPhoto, and I converted 
this to Digikam using a Ruby script (on Github) and a lot of time 
manually correcting stuff. I noticed that some face rectangles were at 
an incorrect position, especially with cropped or rotated images, and a 
lot of labels were simply wrong.

To optimize my future workflow and improve face detection, I would like 
to know a bit more about how Digikams face detection actually works. So 
I have these questions:

1. How big does a face rectangle need to be if I draw one manually? Does 
it need to contain the whole head of the person, or is "eyes + nose + 
mouth" sufficient? Does this size need to be consistent throughout one 
person or does Digikam not care at all?

2. How tolerant is Digikam regarding rotated faces, e.g. when people are 
lying, upside down, looking sideways (= profile view), or looking down 
(skewed view)? Does Digikam regard these faces as "different"? If so, is 
this avoidable? I noticed that after importing a lot of images with one 
person lying down, Digikam would suddenly assign this person to ALL 
photos where ANYbody was lying down.

3. How tolerant is Digikam regarding "decoration" like glasses, scarves, 
half-obscured faces etc? Do such images pollute the learning process?

4. Does Digikam care about exposure, brightness, or sharpness of images? 
Do these images "hurt" the faces learning process?
In other words, can I manually tag images where faces are not detected 
and maybe not even clearly visible, but where I *know* the person is 
there, just for my organization, or will this poison the learning database?

5. How tolearnt is Digikam regarding false positives? If 1 of 100 images 
for a person is incorrect (but confirmed), how much will this interfere 
with the detection of future faces?

I am currently using Digikam 7.9.0 appimage 2022-10-12, on KDE Neon.


Thank you!

-- 
Regards, Jens