State of face recognition in digiKam?

[daj omu]

I totally agree with you that face rec. is not as good as it could be. I
wish that DK some how could identify which images are too blurry og which
face is too low quality to be used as part of its training data. To my
understanding,
DK will perform worse for every bad quality photo added to faces. Maybe
this is only for manuallu added photos?
Also, as people have mentioned before, there should be a way to mark people
without messing up the DB. One can use tags, but not sure if that is a good
solution.

Dajomu


On Mon, 5 Feb 2024 at 11:34, Michael Moore <stuporglue at gmail.com> wrote:

> On Mon, Feb 5, 2024 at 11:20 AM <chimney at thespicers.net> wrote:
>
>> I'm trying to use face recognition in digiKam but am having no success. I
>> have run face detection across my whole collection of around 13,000 photos
>> using the default parameters (70%, not YOLO v3). This completed in a few
>> hours and successfully identified many faces. I then manually identified 14
>> new people from Unknown and assigned several faces to them - at least 5
>> faces for 8 of the people.
>>
>> After this, I have tried using recognise faces, but the results are
>> extremely poor. Each time it runs, it basically assigns all faces to the
>> same one person and the vast majority of these are completely different
>> people that don't resemble them at all. I have not seen it ever assign
>> faces to any other person. The person that it choses has 28 manually
>> assigned faces that are correct.
>>
>> It seems that digiKam simply isn't learning how to recognise faces. This
>> seems to closely match the experience described here:
>> https://bugs.kde.org/show_bug.cgi?id=444160#c10
>>
>> I'm using the latest Windows 8.3.0 snapshot (04/02/2024 10:33).
>>
>> Am I doing something wrong or is this feature simply not working at
>> present?
>>
>
> Hi there,
>
> I can't comment on the official state of face detection, but I can share
> what has been working for me.
>
> I have been running it with 97% Face Accuracy. I have done this for 3
> reasons:
>
> 1. My initial run at the default accuracy produced poor results, so I
> upped the accuracy
> 2. I assume that as I train more faces, the model will improve - 97%
> accuracy generates mostly correct hits. It also has some wrong results,
> which I tag with the correct person instead. This level of correction, I am
> hoping, provides a balance in my work flow to eventually tag everyone.
> 3. I have about 200,000 photos, at 99% accuracy it find few enough results
> that I can review them in a few sessions. Then I run it again and have
> another batch to review.
>
> I also hope for improvements in the process, but with the high accuracy
> setting it seems to do good enough for me for now.
>
> I'm running on Debian Linux with version 8.1.0 and MySQL as my database.
>
> --
> Michael Moore
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.kde.org/pipermail/digikam-users/attachments/20240206/6237fa6d/attachment.htm>