Bug 470018

Summary:	Search for similar faces by previous example selection
Product:	[Applications] digikam	Reporter:	Andy <zunar>
Component:	Faces-Recognition	Assignee:	Michael Miller <michael_miller>
Status:	RESOLVED FIXED
Severity:	wishlist	CC:	caulier.gilles, chrisc.gigamail, fabbriattalo4, michael_miller, quochungtran1999
Priority:	NOR
Version First Reported In:	8.0.0
Target Milestone:	---
Platform:	Microsoft Windows
OS:	Microsoft Windows
Latest Commit:		Version Fixed/Implemented In:	8.5.0
Sentry Crash Report:

Description Andy 2023-05-19 20:35:02 UTC

Hello, 
my facial recognition has a very high error rate. I've already written a lot about this.
At the moment I have over 20000 unconfirmed and 30000 unknown to work through one by one.
My idea would be that I mark five to ten matching faces under "unconfirmed" or "unknown" and then digikam groups new suggestions for me from all other unknown or unconfirmed faces that show an actual similarity. Maybe that would also work with a selection of the already confirmed faces.
When you're happy with the result, mark the first and last faces and confirm them all at once.

Best Regards
Andy

Comment 1 Chris 2023-10-17 08:42:48 UTC

(In reply to Andy from comment #0)
> Hello, 
> my facial recognition has a very high error rate. I've already written a lot
> about this.
> At the moment I have over 20000 unconfirmed and 30000 unknown to work
> through one by one.
> My idea would be that I mark five to ten matching faces under "unconfirmed"
> or "unknown" and then digikam groups new suggestions for me from all other
> unknown or unconfirmed faces that show an actual similarity. Maybe that
> would also work with a selection of the already confirmed faces.
> When you're happy with the result, mark the first and last faces and confirm
> them all at once.
> 
> Best Regards
> Andy

I think the grouping under "Unconfirmed" is actually the automatic version of half of what you suggest. The problem is that the suggestion is probably wrong, since you need to manually select some (5-10 matching faces) and tell the algorhithm that "this is the same person, please find more like it."

In another bug-report / suggestion I saw a mention that the last 100 face tags of a particular person are used by the algorithm to recognize people. Another way to phrase your idea, could be "allow the user to manually select 5-100 face regions of a particular person and make the algorithm use only those".

It's not exactly the same, but it's all about helping the recognition algorithm use the best data. Maybe? :)

Comment 2 caulier.gilles 2024-10-08 20:41:56 UTC

*** Bug 454938 has been marked as a duplicate of this bug. ***