[digiKam-users] GSOC 2017 Project 4 : digiKam Face Management improvements added to DK6?

[Gilles Caulier]

2018-08-24 8:38 GMT+02:00 NeiNei <neinei at gmx.net>:

> Dear Gilles and Stefan,
> thanks for the further clarification face management improvements (AI,
> deep learning).
>
> @ Gilles: Could you estimate whether the still necessary code finalization
> and further improvement will make it in DigiKam 6 or a later 6.x version,
> so that we could use the face tagging improvements in production?
>

Wel at least Few weeks of full time to finalize the code, hack , and test.
The last stage is the most important, as we need to be sure that no side
effects are injected in the rest of the program.

Typically, the recognition data are in a separated database. So the errors
will be limited. I think more about time computation, latency, crashes, etc.

I'm not sure also, if the data stored in database are well managed. The
code is complex and need investigations.

Another question is about the code inside digikam core. i'm sure that we
have too much duplicate and DLib shared library can be used as well in
place. We don't have dependency to DLib. We take the necessery classes for
testing quickly. As the code as good results in recognition, this point
must be changed). In other words, the recognition with AI work but severals
optimisations in source code need to be done.

The good point is that we have now 4 algorithms to process recognition. We
have done comparisons, and Deep learning was the best.

But the problem is : it's slow. AI code in DK core is a demonstrator, we
cannot use it in production.

This is why we don't turn on this feature for the moment. We need help to
investigate. The human resources in DK is limited and this task require at
least one developer to work on it.

We have more important priorities for the moment. So this feature is
delayed, excepted if someone in this room is interested to work on it of
course. You are welcome.

Best

Giles Caulier
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.kde.org/pipermail/digikam-users/attachments/20180824/3fa799e9/attachment.html>