How to improve face recognition process?

[mastodon]

I have about 40.000 pices of family photos (about 10 people) mixed with my collegues and friends (100 people). Unfortanately FR is not so good as I expected. Even faces from front are not detected as a face or are not recognised. What is a good workflow to get the best FR?

I have found, that some people have many-many trained faces (even 500). I have learnt, that is not good for the result of FR.
How can I select the best ones? I use person manager with date of birth. How to use this for selecting enough faces trhouzifetime

Second. Why is that, if I choose the option "Optimize for images with small faces", than sometimes it does not detect normal sized faces?

[Mario]

The face recognition in IMatch is at the height of current research and there is no way to improve it further. Maybe in a few years.
From what users tell me, the face recognition in IMatch works as good, if not better, than in most other software which does FR.

The face recognition usually just works, right out of the box.
How did you use it? Did you process entire folders with hundreds of persons mixed? Manually in the Viewer? The entire database in one go?

500 trained faces is way too much. 2 to 10 faces per person, at different ages is more than sufficient.
IMatch only auto trains the first 10 faces when you confirm faces - which is usually all that is needed.

I wonder how some persons can end up with 500 trained faces.
Use the Face Manager to untrain them all. Go go "Trained", press Ctrl+A to select all trained faces and then T to toggle the trained state.
Then train 3 to 5 faces, at different angles and ages for each person. There are no "best" faces. Use frontal head photos for a start.
If the age of the person varies a lot (and you have entered the day of birth for the person, which is important), train faces at different ages, train 5 to 10 faces in that case.

Over-training like this can badly impact the face recognition. If you have multiple persons for which you trained hundreds of faces each, basically all persons in the database will look alike...

Optimize for images with small faces just uses a larger image for face detection. IMatch cannot detect faces smaller than 40 to 80 pixel, or thereabouts.
This does not impact the detection of other faces, it just allows IMatch to find smaller faces better.

[mastodon]

I don't know how I managed to reach nearly 500 trained faces. I run FR on folders with about 50-250 pictures at once. Than I confirm faces, and make a lot manual face annotations. Actually, I don't understand, where trained faces come from! There must be a failure in my workflow somewhere.
Thanks, I will begin it from the scatch as you pointed out.

[Tveloso]

I don't know how I managed to reach nearly 500 trained faces...

Perhaps this is a remnant of the very early versions of IMatch 2020 (before IMatch capped the auto-trained faces to the first 10 confirmed by the user - and before the Face Manager was introduced, which allows the user to explicitly train selected faces for a person)?...

In those early versions I was having issues with my wife and daughter not being recognized as themselves, and it was in part because of the high number of trained faces that I also had for them, and also due to my having rejected suggestions by the IMatch AI, of my wife for my daughter, and vise versa (my wife and daughter do resemble each other).

Another user pointed out that rejecting faces that the IMatch AI had gotten wrong (because the faces did resemble the person that they were incorrectly assigned to), might then prevent the AI from subsequently selecting the correct face for the person (since it now resembles a rejected face for that person).

I don't know if the "rejected faces issue" is still a consideration, but I have avoided rejecting faces, and IMatch now almost always recognizes all of my family correctly (including my wife and daughter)...I almost take it for granted that the assigned faces are all correct!