Face Recognition: Persons not recognized anymore

[abgestumpft]

Hi,

I'm working on Face assignments at the moment and found that some persons are not recognized at all anymore.
Especially my kids and wife where I have a lot of trained faces.

When I started I went trough a folder in the viewer and confirmed or changed names of Persons there, which everytime generated a trained face. E.g. for my kids / wife I have now around 200-300 trained faces each.

As said, it looks like none of my kids are recognized anymore. E.g. also if I train the face on the RAW (front facing), the corresponding JPG will still have a "?". Also when manually deleting the face annotation and rescan with F4 for new faces the JPG still displays the "?"

What could lead to this (person not recognized anymore)?
1. Too many trained faces for this person?
2. Too many "Not this person" actions done -> not sure if this is saved in any way
3. Too many trained faces in sum: Persons (I have 44 at the moment)...

[Jingo]

I'd recommend just deleting the training faces and retraining the person on about 5 faces... I too had this issue in the beginning (one person had about 100 faces in their training... almost all the same) but Face Detection has evolved even from the first release.. so it would make sense to remove the trained faces and start again.

[abgestumpft]

Thanks for your reply.
I'm new to iMatch and using it starting with Version 2020.3.2 (I think this was the 30-days trial).

What I did now:
1. In people editor I removed ALL trained faces of my Wife
2. I took four images I shot in a row where she was looking similar
3. I trained the first face in Viewer, Closed the viewer and checked that this new face was now visible in trained faces -> Yes
4. For the other three I added the face annotation with "F4" in the viewer -> wrong person suggested
5. I trained two more faces in the viewer -> now three of those four are in the trained faces -> still wrong person suggested on the forth image
6. On the fourth image, when I delete the face annotation and create it again -> still wrong person suggested

[Wolfgang]

I assume in your situation that other persons with a too high number of trained faces cause the suggestion of a wrong person.

I assume that too many trained faces not only interfere with the face recognition algorithm for a specific person but disturb the recognition process for "all" faces.

In my situation, I have about 400 persons with some of them more than 1000 trained faces (cause: import from face annotations in IMatch 2019, initially coming from Picasa).
I now work my way through all these persons and reduce the number of trained faces to about 5-15 (relatively easy, as explained in my posting today on the thread "Face detection: Confirmed faces vs trained faces").
I expect good face recognition with IMatch 2020 only after no person has more than this recommended number of trained faces.

If I'm right, any confirmation leading to additional trained faces (like confirmation in the Viewer) could easily push some persons over a critical limit and a new action to reduce the number of trained faces would be required in order to improve the overall face recognition. I assume the Face Reviewer (Mario is working on) will help in the future.

[abgestumpft]

I assume in your situation that other persons with a too high number of trained faces cause the suggestion of a wrong person.

My two kids have the highest number of trained faces (ca. 200 - 250 each) and have like 0% hit-rate at the moment.
When there is a picture of one of my two kids, usually no face is suggested at all and only the "?" displayed.
Just was working on some pictures of my brother and his wife -> here almost 100% hitrate...

I assume that too many trained faces not only interfere with the face recognition algorithm for a specific person but disturb the recognition process for "all" faces.

That might be the case. I'm also doing a lot of "Not this person", "Reject Person", Deleting Face Annotations, copying face annotations, ...
Might be that this also has a negative impact on face recognition

In my situation, I have about 400 persons with some of them more than 1000 trained faces (cause: import from face annotations in IMatch 2019, initially coming from Picasa).
I now work my way through all these persons and reduce the number of trained faces to about 5-15 (relatively easy, as explained in my posting today on the thread "Face detection: Confirmed faces vs trained faces").
I expect good face recognition with IMatch 2020 only after no person has more than this recommended number of trained faces.

I'm trying this at the moment (reducing the trained faces massivley for each person), but so far the hitrate for my two kids is still close to 0%.

If I'm right, any confirmation leading to additional trained faces (like confirmation in the Viewer) could easily push some persons over a critical limit and a new action to reduce the number of trained faces would be required in order to improve the overall face recognition. I assume the Face Reviewer (Mario is working on) will help in the future.

This is kind of my problem at the moment -> my kids/wife are not detected at all anymore -> I have to always put the name in via the viewer -> number of trained faces grow for them...