Face recognition and keywords

[ben]

Hello,

would someone please help and explain to me, how iMatch handles the following situation:

I have defined person "mike" and added the keyword "persons|mike" in the person editor.
I have a picture with several persons, one of them is mike.
I manually assigned the keywork persons|mike.
I now run face recognition but none of the persons is recognized as mike, but as other persons.

Does iMatch remove my previously assigned keyword persons|mike?
If that's the case, how can i be sure, that my manually assigned keywords are still there, after i run face recognition?

Thanks for your help
Ben

[Mario]

Have you actually tried this?
What was the result?
This should be easy for you to produce and observe.

Please understand that I cannot fathom or prepare for any possible fringe case. Doing so would spoil the software for the majority of users.
Like, mixing manually assigned keywords with the same person-based keywords in conjunction with more or less accurate detection rates. I doubt many users would ever try that.
Maybe keep manually assigned keywords and person-based keywords separate if you anticipate any troubles in that area?

[ben]

Have you actually tried this?
What was the result?

Yes, of course.
The result is as i described it, maybe explained it not well ;-)
iMatch removes my previously manually assigned person-keywords.
But it shouldn't since i know they are right.

So, i wanted to give the iMatch face recognition a second (third) chance.
But it seems i get stuck at the same point again.

So the obstacle i see for me is:
Just relying on the face recognition doesn't work, since there are faces that are not recognized or not even detected.
Now there is the function "link persons", but that's way less comfortable to use than the dockable keyword panel which provides access to keywords in many ways.

Ok, so it seems that's the way it works.

Does anyone else use manually assigned keywords in combination with face recognition?

[Mario]

iMatch removes my previously manually assigned person-keywords.
But it shouldn't since i know they are right.

This is what I said. IMatch cannot know that you mix person keywords with keywords you manually assign.
If the AI swaps a person (or you change the person assignment), IMatch removes the keywords linked to the previous person and assigns the keywords now linked to the face.
There is no way to tell "But if the user has manually assigned one or more of the person-linked keywords, don't change anything".

If you mix manually assigned keywords and automatic person keywords, this will cause issues.
Either maintain separate keywords or manually correct wrong person assignments, which also corrects the keywords.

[ben]

If you mix manually assigned keywords and automatic person keywords, this will cause issues.
Either maintain separate keywords or manually correct wrong person assignments, which also corrects the keywords.

I tried maintaining different keywords (persons_AI|<name> vs. persons|<name>) but i failed to create data driven categories to merge these two.
I also tried correcting wrong assignments and linking people but i am way slower than using the keyword panel (which offers so many possibilities which the link/unlink window doesn't).

Don't get me wrong. I don't want to be nagging, i am just a bit frustrated, that i cannot get the face recognition to work for me. And i don't understand that others are not facing the same problems. Wrong face recognitions and not detected persons should be common, even if the iMatch recognition rate is great.

So have a nice sunday
Ben

[Mario]

but i am way slower than using the keyword panel

The Keyword Panel is exactly designed for this. Adding keywords to many images very fast, with checks, common vocabulary etc. Favorites are even faster (if your frequently used keyword vocabulary is small enough).
Using people-keywords is comfortable and works very well if the recognition rate is good in general.

The IMatch face recognition is very good, but it cannot be 100% correct.
I've made the "training" faces used public so users can correct wrongly trained faces or swap with "better" faces. This often further improves recognition rate for edge cases. Did you try that?

All face-related features combined, this is what is currently technically possible. Or at least what I can do. Looking at the recognition rate of other software, IMatch is pretty good 

I tried maintaining different keywords (persons_AI|<name> vs. persons|<name>) but i failed to create data driven categories to merge these two.

Have you tried using the replace function to replace "persons_AI" with "persons" to merge? Or maybe use a variable as the source for your data-driven category, for more complex replacements or folding if needed.

When I use a variable like {File.MD.XMP::Lightroom\hierarchicalSubject\HierarchicalSubject\0|replace:_AI==} with detect hierarchies via the | character for a data-driven category, IMatch folds the keywords

Persons_AI|Frank
Persons_AI|Lilly
Persons|Tom
Persons|Laura

into



which should do the trick for your situation. This will not impact the keywords actually assigned and used in @Keywords, but it allows you to fold different hierarchies of keywords in a custom data-driven category.