Re-train facial recognition?

[Aaron de Bruyn]

Hey everyone,

I've been playing with the face recognition in DigiKam for a few months now trying to sort through about 500,000 photos, and I've run into a few problems.

I'm making a few assumptions here, so please let me know if I'm way off-base.

I started out by tagging a few hundred photos of various family members (let's just say Person 'A', 'B', 'C', etc...).
After I did that, I ran the recognition at 90% and it managed to get a few images that were nearly duplicates of the ones I tagged. I confirmed them.

Then I dropped it down to 80% and a bunch were detected incorrectly.

While going through several thousand images...at one point I accidentally tagged 20 or 30 images of Person A as Person B.

I noticed there's no "Undo", and that's understandable based on my knowledge of how training databases work.
Anyways, I went in to Person B, found the images of Person A and corrected them to be Person A.

But from that point on, I appear to be doomed.

The system still frequently detects Person A as Person B, and I've made the mistake a few more times while classifying images.

Yesterday I went through every single face tag and corrected probably 200 images in total that were confirmed incorrectly.

Anyways, I re-ran the recognition tool, and it's still guessing people incorrectly...

My guess is the training database "learned" my mistakes and correcting them doesn't really undo the damage I initially did.

Is there a way to "reset" the facial recognition database and re-train it from scratch by saying "The images I have already tagged as Person A, Person B, etc... are correct"?

Is that what "Clear and rebuild all training data" does in Maintenance? I'm hesitant to try that option as it will probably take a solid day or two to run and I don't want to lose all the manual tagging I've done.

Thanks!

-A
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.kde.org/pipermail/digikam-users/attachments/20230407/6170edbc/attachment.htm>