[digiKam-users] Face Recognition Workflow

Thu Aug 19 13:27:50 BST 2021

I have the Same issue here and would be interested in some guideline.

I have seen there is one project (suggestion?) of the Google summer of code "Project: Improve Image Quality Sorter algorithms" (https://community.kde.org/GSoC/2021/Ideas#digiKam).
Can this be used to "filter" the faces used for training to overcome this?

Cheers,
Thomas Beckler

________________________________
Von: Digikam-users <digikam-users-bounces at kde.org> im Auftrag von digikam-users-request at kde.org <digikam-users-request at kde.org>
Gesendet: Donnerstag, 19. August 2021, 13:00
An: digikam-users at kde.org
Betreff: Digikam-users Digest, Vol 195, Issue 20

Send Digikam-users mailing list submissions to
        digikam-users at kde.org

To subscribe or unsubscribe via the World Wide Web, visit
        https://mail.kde.org/mailman/listinfo/digikam-users
or, via email, send a message with subject or body 'help' to
        digikam-users-request at kde.org

You can reach the person managing the list at
        digikam-users-owner at kde.org

When replying, please edit your Subject line so it is more specific
than "Re: Contents of Digikam-users digest..."

Today's Topics:

   1.  Face Recognition Workflow (Rhys Tyers)
   2. Re:  Face Recognition Workflow (Travis Kelley)

----------------------------------------------------------------------

Message: 1
Date: Wed, 18 Aug 2021 18:26:59 +0100
From: Rhys Tyers <rhystyers1 at gmail.com>
To: digikam-users at kde.org
Subject: [digiKam-users] Face Recognition Workflow
Message-ID:
        <CAEg=DQH=-dJfOY-EYii0DvuM4CPdgz_69Gmw5tX0LZC1nVBzuw at mail.gmail.com>
Content-Type: text/plain; charset="utf-8"

Hello,

What is the correct way to train the face detection model when many of
the faces are poor quality images?

I have tens of thousands of photos which include hundreds of people many
hundreds of times each. Many of their faces in these photos are quite poor
quality (obscured, dark, far away).

When I start recognition it is often quite poor but once I have a few high
quality faces in the model it seems to work fairly well. However as I
accept more of the poor quality faces I find the recognition gets worse (I
assume because dark, far way, poor quality faces look very similar). People
for whom I've tagged a lot of poor quality images just start getting any
face suggested for them.

What is the intended workflow? The application prompts you to accept
recognised faces as face tags (and then they will get used in the
training), so that's what I've been doing.

But then there is this in the manual:* In case of unsatisfying results it
might be helpful to use Clear and rebuild all training data. One reason can
be that there are too many face tags assigned to a person which shows this
person in a way that doesn't really help the search algorithm, e.g. with
sunglasses, blurred, unusual colors, carnival make up, dark shaded areas in
the face, baby/kid/adult photographs mixed.*

So that implies that I should only add high quality faces to the face tags
and perhaps tag poor quality ones differently? Is there some way to mark a
face as "poor quality" so it is not used for training?

Thanks for any help.

Rhys
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.kde.org/pipermail/digikam-users/attachments/20210818/4a6516ed/attachment-0001.htm>

------------------------------

Message: 2
Date: Wed, 18 Aug 2021 13:33:28 -0400
From: Travis Kelley <rhatguy at gmail.com>
To: digiKam - Home Manage your photographs as a professional with the
        power of open source <digikam-users at kde.org>
Subject: Re: [digiKam-users] Face Recognition Workflow
Message-ID:
        <CAC0Og8nAC_vREX3ygL9WUL4UBedLiP5qMtds0dqyDjPXJ2f9JQ at mail.gmail.com>
Content-Type: text/plain; charset="utf-8"

I'll plus one this issue.  I have the same problem, where it seems that the
more you tag faces the worst the algorithms get.  As Rhys mentioned, not
sure if the answer is to mark faces as bad or whether it would be better to
be be able to somehow specify which faces to use for training.  Perhaps
there is a programmatic way to detect the quality of tagged faces and
select the "best" images for training?  I believe there is a limit of
something like the last ~100 tagged faces used for training today right?

On Wed, Aug 18, 2021 at 1:27 PM Rhys Tyers <rhystyers1 at gmail.com> wrote:

> Hello,
>
> What is the correct way to train the face detection model when many of
> the faces are poor quality images?
>
> I have tens of thousands of photos which include hundreds of people many
> hundreds of times each. Many of their faces in these photos are quite poor
> quality (obscured, dark, far away).
>
> When I start recognition it is often quite poor but once I have a few high
> quality faces in the model it seems to work fairly well. However as I
> accept more of the poor quality faces I find the recognition gets worse (I
> assume because dark, far way, poor quality faces look very similar). People
> for whom I've tagged a lot of poor quality images just start getting any
> face suggested for them.
>
> What is the intended workflow? The application prompts you to accept
> recognised faces as face tags (and then they will get used in the
> training), so that's what I've been doing.
>
> But then there is this in the manual:* In case of unsatisfying results it
> might be helpful to use Clear and rebuild all training data. One reason can
> be that there are too many face tags assigned to a person which shows this
> person in a way that doesn't really help the search algorithm, e.g. with
> sunglasses, blurred, unusual colors, carnival make up, dark shaded areas in
> the face, baby/kid/adult photographs mixed.*
>
> So that implies that I should only add high quality faces to the face tags
> and perhaps tag poor quality ones differently? Is there some way to mark a
> face as "poor quality" so it is not used for training?
>
> Thanks for any help.
>
> Rhys
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.kde.org/pipermail/digikam-users/attachments/20210818/536e4852/attachment-0001.htm>

------------------------------

Subject: Digest Footer

_______________________________________________
Digikam-users mailing list
Digikam-users at kde.org
https://mail.kde.org/mailman/listinfo/digikam-users

------------------------------

End of Digikam-users Digest, Vol 195, Issue 20
**********************************************

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.kde.org/pipermail/digikam-users/attachments/20210819/19bc328a/attachment.htm>