Thinking about auto-tags (was: Re: Feature request for Digikam)

[Kjetil Kjernsmo]

Hi all!

I found auto-tagging very interesting, but departing from the original 
subject, so I figured it was best changing the subject:

On mandag 26. februar 2024 16:48:49 CET Gilles Caulier wrote:
> About the auto-tags project, take a look to the student blog :
> 
> https://community.kde.org/GSoc/2023/StatusReports/QuocHungTran#Add_Automatic
> _Tags_Assignment_Tools_and_Improve_Face_Recognition_Engine_for_digiKam

I've spent a lot of time on tagging over the years, I currently have some 
17000 tagged images in my library, and I would be happy to help training 
models with these images and tags, if the original images didn't need to go 
anywhere, in other words, if federated training was supported, and if the 
model had an acceptable license (I would suppose that's taken of since Digikam 
uses it).

Of course, I also have tags of family members that I wouldn't want to expose, 
so Digikam would need to separate between tags for public consumption and my 
private tags. However, my private tags would also be useful. I've used face 
tagging a bit, but it hasn't replaced my manual tagging of people, simply 
because I want to know who's in the picture, also when the face is invisible. 
I can usually make this out manually based on what clothes they are wearing, 
and so it seems to me that this solution could enable person recognition 
beyond facial recognition. If this enabled me to just tag a few images and 
then have a model that figures it out based on recognition of something like 
clothing, it would save me a lot of time ;-)

This direction also seems to create demand for offloading processing to other 
machines, as ML tends to be pretty heavy. My laptop is light and silent, but I 
have a bigger box in my basement. Its disks actually hosts my pictures, I use 
NFS and a remote MySQL database, so if I could run the AI stuff on that box 
directly from my laptop, that would be very nice. Any chance that for example 
the WorkerObject could be extended to run on a remote machine, given, say a 
shared database and an NFS or SMB share?

So, it seems to be a lot of potential here, in terms of crowd sourcing model 
training in a privacy-preserving manner so that the capabilities of Digikam 
improves for everyone.

Kind regards,

Kjetil