Face recognition suggested improvement.

[Travis Kelley]

There are ways to programmatically detect blurry images also.  Maybe there
could be some programmatic way to have blurry images affect the model less
(or not at all).  Other apps don't seem to be affected in this way.  Mylio
for instance doesn't degrade when tagging blurry images.

On Tue, Feb 20, 2024 at 9:17 AM Nic Garton <ngarton at hotmail.com> wrote:

> Hi
>
> Digikam - wow! I love it.
>
> Now, to possible improvements.
> As has been pointed out by many, face recognition tends to fail to almost
> useless once you train the engine with some blurred or obscured faces.
> I have got round this by emptying the face trainings, and starting over
> with some clear faces.
> But it quickly degrades again.
>
> To empty the face trainings, and start over, I have been running the
> following SQL commands, in my favorite db client (Dbeaver). Obviously need
> external SQL DB for this.
>
> delete from FaceMatrices;
> delete from Identities;
> delete from IdentityAttributes;
> delete from ImageTagProperties where property = "autodetectedPerson";
>
>
> (I keep everything I want to keep in sidecar .xmp files, so I don't worry
> about messing the DB itself).
> So, my suggestion to make this work good, and it is SO CLOSE to working
> good, is to have a gui interface to be able to select which faces are used
> for each identity.
> Then if you "accidentally" add a blurry one, it can just be removed
> through the gui. So each identify would keep a set of say, 5, really good
> faces images.
> Then recognizing a face to this identity would NOT automatically add it to
> the trainings for that face ("FaceMatrices" table).
> In short, much more user control over which faces make the training set,
> and which don't.
>
> I have no idea how much effort it would take to add this interface, but
> based on my tests, it would solve many of the problems, and make this an
> even more fantastic product!
>
> Thanks in advance!
>
> Nic
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.kde.org/pipermail/digikam-users/attachments/20240220/f292b64e/attachment.htm>