https://bugs.kde.org/show_bug.cgi?id=435863

--- Comment #7 from [email protected] ---
Dear Gilles, dear all,

meanwhile I have successfully transited my 6TB image archive from digikam 6 to
version 7. I am not on 7.5 yet. Currently running 7.3.0 app-image under ubuntu.
It may still be useful to some, if I provide some feedback here, especially as
I had hit a pothole on my first attempt and I can also mention a few
frustrating experiences on this second succesfull try.
But first of all I want to make clear that I am very happy with what I have
running now. Digikam for me is the workhorse I am using all day. Once the
transition of the vast amount of faces to the new face recognition is done,
everything is just fine. The transition is a one time effort and probably I did
not have the optimal strategy for that. But I succeeded.
Now, where did I encounter the traps? To play it safe, I tested on a second
computer rather than taking the risk to interrupt my daily work. The tests with
a subset of data worked nicely but I learned that the detection of faces uses a
huge amount of computing. As I was close to upgrade my system anyway, I decided
to buy a new computer first. What I finally use is an I7 8-core with 32 GB RAM,
500 GB SSD and 12 TB hard disk. Images reside on the harddisk, digikamdb on the
SSD. It took me more than a week to re-scan all the faces. As many of my images
contained pre-tagged faces originating from picasa or manually tagged within
digikam 6, I wanted  to start with a clean data base and expected digikam 7 to
learn from the tags within the image files. 
I would load chunks of images for the computer to stay busy detecting faces
over night and next morning I would assign names where needed, confirm
digikam's suggestion or correct them in order for the AI to learn the faces. 
For the first two or three days that would work fine. However, digikam is
simply "too good" in finding face areas. This leads to an enormous amount of
unknown blurry faces in the "unknown" Folder. There is a need to remove those
by marking "ignore" because with large amount of data the user has no chance to
get an overview. It would be very useful if there was a parameter to specify
the sensitivity with respect to a minimum amount of pixel a face should have
and with respect to what is an acceptable blurriness. That simply to reduce the
amount of people found somewhere in the background.
The second trap is somehow related to this. After a while, the face recognition
started to produce more and more bogus-results and it took me a long time to
come up with an idea why this happens. I found a valuable hint somewehre else
in the forum. According to that post, digicam only uses a limited number of
samples of a person's face to compare with an unknown face. In my case with
many pre-tagged faces this leads to digikam un-learn faces. I would like to
illustrate this with an example: Every month I shoot a Jazzband performing. The
person in the foreground was usually pre-tagged. So when the image is scanned
the dominant high resolution foreground faces are immediately added to the data
base. However many images will have the drummer in the background somewhat
small and out-of-focus. Digikam will surely find all these faces and present
them for confirmation. For the photographer it is easy to recognize the
drummer, so click "confirm" or assign his name, if he shows up under a wrong
name. However this means that you load all the blurry instances of the given
person and after you add enough, digikam will not use the sharp ones for
identification any more. Boom! It was very frustrating until I changed my
strategy to simply ignore all blurry images even when I could easily say who it
is. I hope this explanation is somewhat correct and helps others to avoid to
repeat my mistake.
What I am missing is a mechanism that would sort or group unknown faces by
similarity. I believe that was something picasa did well. With such a grouping
mechanism it would be possible to keep a large amount of faces in the unknown
folder rather than having to mark them "not-a-face". This way the software
would tell, that you came across the same face year after year but never
bothered to put a name to the person. A nice incentive to do a little research
to find out, who this person is. As I shoot a lot of open air events, this
turned out very valuable in the past. 
I understand that the use-case I illustrated here, is not what digikam was
designed for. and I repeat, once you have worked your way through this initial
transition, everything works just fine since you will typically add new faces
with your new images.
I would like to apologize, if I missed items which are actually already there
while I was too ignorant to find it. I am just a naive user, who tries to get
things done without too much knowledge about the software. As I said before:
probably the bug sits on the chair in front of the keyboard in my case.

-- 
You are receiving this mail because:
You are watching all bug changes.

Reply via email to