Face detection: Confirmed faces vs trained faces

[ubacher]

Should the number of confirmed faces not be the same as the number of trained faces?
I find they are not the same.

[Tveloso]

I found these two statements in the help last night:

When you assign a person to a face via the Person Selector or you confirm a suggestion made by the IMatch AI in the Viewer, this face becomes a trained face for that person.

Do not train too many faces. To many trained faces will cost a lot of performance and may even produce worse results!

That's a very important distinction!

I have seen other posts saying the the more faces of a given person were confirmed, the lower the recognition rate for that person became.  And I too started seeing that for my daughter.  It turns out that I had mostly been assigning or confirming in the Viewer, and creating a large set of Trained Faces for people.

My next task, before I proceed with face recognition will be to whittle down the trained faces to a small representative set for each person.  Thank you so much, Mario, for that feature.

[Mario]

Should the number of confirmed faces not be the same as the number of trained faces?
I find they are not the same.

What are Trained Faces?

When you assign a person to a face via the Person Selector or you confirm a suggestion made by the IMatch AI in the Viewer, this face becomes a trained face for that person.
Batch-confirming faces via the corresponding commands in the File Window does not train these faces.

To many or the wrong trained faces are counter-productive.

[ubacher]

When you click on the smiley icon in the person editor you get trained faces.

When I confirm a face it should be added to these trained faces - according to my understanding.

[Mario]

You confirm how?

[jch2103]

Too many or the wrong trained faces are counter-productive.

In reviewing Trained Faces in the Person Editor, I'm finding that some persons have many more trained faces than others (over 500 in one case), and in these cases I find runs of nearly identical faces (many of these apparently from RAW/JPG pairs). My faces/Persons come from a variety of sources over time, including Picasa, Lightroom and now IMatch import.

Are you suggesting that I should prune these trained faces?

[Mario]

500 faces is way too much.
Select All and un-train (click in the train dialog for instructions).
Then train 10 or 15 faces for the person. More than sufficient.

[jch2103]

How did they all get there? Did IM use all of the previously recognized faces (from Picasa/Lightroom/etc.) and add them to the recognized persons?

[Mario]

May be. I have changed so many things in the past four weeks and users have indexed files with XMP face data in all these weeks with different states of the face engine.
Just select them all, untrain. Then re-train. Should take maybe 60 seconds.

[jch2103]

May be. I have changed so many things in the past four weeks and users have indexed files with XMP face data in all these weeks with different states of the face engine.
Just select them all, untrain. Then re-train. Should take maybe 60 seconds.

Remind me exactly how to do that. Is untraining the same thing as deleting the trained faces?

Thanks.

[claudermilk]

Wrapping my head around this tool. Is training and confirming faces different, or the same thing? I have a big database that I am adding face annotations to and don't want to dig myself a big hole right out of the gate here.

I've run the detect process on the images I've categorized with people in them. Now I have a big batch of images in the Unconfirmed People collection I'm processing. It seemed from the help that I need to slog through those and confirm. Now I'm having second thoughts?

[Mario]

Remind me exactly how to do that. Is untraining the same thing as deleting the trained faces?
Thanks.

Open the Person Editor  and click on the smiley icon.
This shows you all trained faces for the person. To un-train a face, just click on it.
You can also select all faces and un-train them all.

See Reviewing Trained Faces and 'Untrain' Faces

You can then train new faces using the <F2> key and the Person Selector in the Viewer.
Or you just don't un-train all but keep 5 to 10 representative faces (different age, different facial expression).

[Mario]

I've run the detect process on the images I've categorized with people in them. Now I have a big batch of images in the Unconfirmed People collection I'm processing. It seemed from the help that I need to slog through those and confirm. Now I'm having second thoughts?

You need to confirm faces manually in the Viewer or the File Window. The AI can only confirm faces automatically which have a high confidence.
This is a manual process. Tip: Use <Tab> in the Viewer to select the face and press <C> to confirm. Faster than mouse clicking.

I'm currently working on a Face Reviewer feature which did not make it into the initial 2020 version. I had to call it quits and "good enough" at some time. To get the thing out, the money and user feedback in 
This feature will streamline the confirm, train and reject tasks massively.
Maybe you'll just wait for a couple of weeks before confirming all faces manually.
On the other hand, I have done this to tens of thousands of images during development and testing, and it is not too much work.

[Wolfgang]

Hello,

also back after a long time. Excited as well about the new possibilities with face recognition. Thanks for all the work!

I'm at a similar stage on my side, having imported from Picasa about 400 persons, all of them have now all faces marked as trained.

Here my actual workflow how to reduce the number of trained faces:

I go to each person which has more than about 10 trained faces, click on the smiley icon, this shows all trained faces.
I click know on all faces which I consider to be the best for training (different age, different facial expression, as Mario wrote).
By clicking on these "best" faces, they get marked "fade away" as untrained.
Now I use the command Ctrl + I to invert my selection. As a result all "best" faces are now shown as trained, all remaining faces as untrained.
I save my new selection with "ok".

One other item is that context:
In his last reply to this subject Mario wrote "You need to confirm faces manually in the Viewer or the File Window".

Just be careful with the difference between these two possibilities. If you confirm faces manually in the Viewer, these faces will end up as trained faces for that person, as explained in the Help. So to avoid a too large number of trained faces, I suggest to use the File Window and not the Viewer to confirm faces manually, if more trained faces are not necessary or beneficial.

[Mario]

You can easily un-train unwanted faces via the Person Editor. Click on the smiley icon at the top and then on the ? for usage hints.

[Tveloso]

Now I use the command Ctrl + I to invert my selection. As a result all "best" faces are now shown as trained, all remaining faces as untrained.

Thank you so much for pointing that out Wolfgang.  I can't believe that I missed that option!

I too have been going through and selecting only the "best" Trained Faces for each person, but have been clicking on all the ones to remove...(many more clicks that the way you have outlined).

I have also considered whether I should remove all trained faces for a person, and then also the Face Annotations from which they were created (since those Annotations were not originally created by the IMatch AI), and re-create them in IMatch.  Not all - just those that would be used for the Trained Faces.

Mario, do you think there would be any benefit in doing that?...(I now you mentioned previously that Annotations created by Apple Devices in particular, were somewhat different from what IMatch creates).

[claudermilk]

Here my actual workflow how to reduce the number of trained faces:

I go to each person which has more than about 10 trained faces, click on the smiley icon, this shows all trained faces.
I click know on all faces which I consider to be the best for training (different age, different facial expression, as Mario wrote).
By clicking on these "best" faces, they get marked "fade away" as untrained.
Now I use the command Ctrl + I to invert my selection. As a result all "best" faces are now shown as trained, all remaining faces as untrained.
I save my new selection with "ok".

One other item is that context:
In his last reply to this subject Mario wrote "You need to confirm faces manually in the Viewer or the File Window".

Just be careful with the difference between these two possibilities. If you confirm faces manually in the Viewer, these faces will end up as trained faces for that person, as explained in the Help. So to avoid a too large number of trained faces, I suggest to use the File Window and not the Viewer to confirm faces manually, if more trained faces are not necessary or beneficial.

Thanks for that tip, Wolfgang. That's a couple of good ideas I'll keep in mind.

You need to confirm faces manually in the Viewer or the File Window. The AI can only confirm faces automatically which have a high confidence.
This is a manual process. Tip: Use <Tab> in the Viewer to select the face and press <C> to confirm. Faster than mouse clicking.

I'm currently working on a Face Reviewer feature which did not make it into the initial 2020 version. I had to call it quits and "good enough" at some time. To get the thing out, the money and user feedback in 
This feature will streamline the confirm, train and reject tasks massively.
Maybe you'll just wait for a couple of weeks before confirming all faces manually.
On the other hand, I have done this to tens of thousands of images during development and testing, and it is not too much work.

OK, thanks Mario. At some point you have to release. There's always something more to add. I'm glad you got the main part out--it's going to be super helpful.

I have figured out the Tab-C/I-left arrow key sequence to quickly move through images.

I may still be in the midst of this project when you release. I have about 30k images to review.    I also realized I want to take advantage of both the auto-keywording when confirming a face annotation, and the general auto-tagger feature. THAT showed me I need to review & update my keywords and thesaurus. So I hit pause on the faces and am updating the keywording first.