Face Detection in DK 8.5 - does it learn?

[Michael Miller]

Hi Steve,
For face detection (is this a face?), the model is fixed. No additional training is done by the end user's actions.

There are even more improvements coming in 8.6.0 for both face detection and recognition.

Cheers,
Mike



Sent from my T-Mobile 5G Device


-------- Original message --------
From: Steven Robbins <steve at sumost.ca>
Date: 1/12/25 4:10 PM (GMT-05:00)
To: digikam-users at kde.org
Subject: Face Detection in DK 8.5 - does it learn?

First, I will say again that I am over the moon with the improvements in
version 8.5.  Face detection & recognition is orders of magnitude better than
in the past.

I just have a small question to clarify my understanding on the detection part
-- i.e. when the face rectangles are created.

I am using the YuNet model, with default parameters (Accuracy = 7, Face Size =
Large).  I find a fair number of "non face" regions identified.  The manual
advises to  use the "Remove this face region" option.  Which is fine: the
rectangle is removed and I'll never be bothered again for this picture (unless
I use options to re-scan the picture).

My query is whether removing a face region will affect future model accuracy in
any way.  For example, there are often many similar pictures of the same
scene.  If a particular tree trunk is detected as a face, it shows up in all
pictures.  If I were to remove the region, is there any update to the model
that reduces the chance that the particular tree trunk will be detected as a
face in a future image of the same scene?

Thanks,
-Steve
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.kde.org/pipermail/digikam-users/attachments/20250112/a27d533f/attachment-0001.htm>