Bug 510254

Summary:	Recognition information per year
Product:	[Applications] digikam	Reporter:	Marcos Dione <mdione>
Component:	Faces-Recognition	Assignee:	Digikam Developers <digikam-bugs-null>
Status:	REPORTED ---
Severity:	wishlist	CC:	metzpinguin
Priority:	NOR
Version First Reported In:	8.7.0
Target Milestone:	---
Platform:	Debian unstable
OS:	Linux
Latest Commit:		Version Fixed/Implemented In:
Sentry Crash Report:

Description Marcos Dione 2025-10-05 07:50:14 UTC

I have been taking pictures for more than 20 years. Since then, my friends and me have grown grey hair, my parents have considerably aged and my kid were born and grew. I have the impression this has an impact on facial recognition. The most evident case is that since I have more pics from my first born, most baby pictures are associated with them. Also I don't see my parents that often (less than once per year), and I have the impression the algos struggle to find the changes for such long periods.

Comment 1 Maik Qualmann 2025-10-05 08:32:47 UTC

Which digiKam version are you currently using on Debian?

Maik

Comment 2 Marcos Dione 2025-10-05 08:39:52 UTC

8.7.0 from Debian Sid.

Comment 3 Maik Qualmann 2025-10-05 08:51:06 UTC

We've actually decided against linking face recognition to age, date of death, or anything like that.

If you retrain people of different ages, it doesn't work?

Maik

Comment 4 Marcos Dione 2025-10-05 09:48:29 UTC

Out of curiosity, is there any record of that discussion? I swear I will not refloat it.

OTOH, am I supposed to retrain from time to time?

Comment 5 Maik Qualmann 2025-10-05 11:17:14 UTC

These are essentially duplicate bug reports:

Bug 468623
Bug 503788

Maik

Comment 6 Marcos Dione 2025-10-05 15:41:25 UTC

Well, https://bugs.kde.org/show_bug.cgi?id=468623 is not what I had in mind. https://bugs.kde.org/show_bug.cgi?id=503788 is closer, but this way it would be more 'natural'. I was thinking about a recognition DB per calendar year. This way not yet born or already dead (or even people who drastically change their appearance) would be supported.

Of course, this would mean that for some people that do not appear so often in photos, training would have to be redone because the other matches would be in other recognition DBs. Maybe you could try to recognize first on the year of the photo and if not a good match, then other years "around".