Looking for suggestions to fix QuMagie face clustering

English Multimedia
1

A few details to better understand the situation:
I have more than 150,000 photos dating back to the early 2000s.
I have more than 5,000 face clusters (groups), of which 332 are named.
On June 14, I noticed that face clustering was stuck at 294,000% and would not stop. The issue was resolved with the help of the support team, but I now suspect that some data may have become corrupted.

Recently, I noticed that one face cluster contains many photos that do not belong there: girls, boys, young people, older people, with no apparent similarity between them. Let’s call this group Joe. Joe is not a family member and should have no more than 100 photos associated with him.

I thought the easiest way to fix the issue would be to delete the Joe cluster and recreate it later. I didn’t want to manually correct 3,000+ photos, many of which contained multiple incorrect recognitions of Joe.

Unfortunately, after deleting the cluster, it did not remain as an “Unknown” group. Instead, it appears to have been merged with another person, let’s call her Marie. I can’t be certain that exactly the same faces were reassigned to Marie, but it certainly looks that way.

I have invested a lot of manual work into organizing these face groups, so I would prefer not to start over from scratch.

Do you think removing photos year by year from the QuMagie library and letting it rebuild the face database could help? I would lose some metadata and manual corrections, but that would still be better than losing everything.

Do you have any other suggestions?

For example, the oldest photo with an incorrect match involving Marie dates back to 2020. Joe is actually in the photo, but it is now labeled as Marie. Before I deleted the Joe cluster, this photo was recognized correctly.

The real mess seems to start in 2022. From that point onward, a large number of different people are being recognized as Marie. I’m considering moving all photos from 2022 onward out of the QuMagie library, manually correcting the older photos involving Joe and Marie, and then gradually adding the newer photos back in batches.

If my understanding of how face recognition works is incorrect, please correct me.

2

Hi @kbazsi this behavior is linked to a known AI loop issue, could you NOT Rebuild or Re-index which may cause further database damage, before the development team confirm your issue?

Could you help to share the version of your QuMagie / Multimedia Console / AI Core with use?

3

Thank you for your reply.
Understood, I’ll wait for further updates.

QuMagie: 2.9.3
Multimedia Console: 2.10.4
AI Core: 3.7.1

4

We’d also like to know whether your face recognition grouping includes any categories you created manually. Thanks as well for your feedback — our internal team will also review possible solutions for the related mechanism.

5

Hi,

I usually assign names to people recognized by QuMagie, but in some cases I have manually tagged people in photos. I don’t recall ever creating a group based solely on manually tagged faces, but I can’t rule out the possibility.

6

After checking internally, we’ll still need to ask you to open a support ticket so our team can look into this further based on your actual environment. Thanks!

7

–English translation added by Dolbyman–

May I jump in here as well?
I’m having the same problem on my TS-451+ NAS system.

Facial recognition used to work really well. Lately, however, the AI has been mixing up the recognized faces in newer photos.

I currently have 151,000 photos processed; recognition seems to have worked fine for older images, but the matching is incorrect for many newer ones. I haven’t yet figured out the import date from which the errors started occurring.

AI-Core version 3.8.1
Multimedia Console version 2.11.1.

No third-party apps installed—only original QNAP apps.

8

According to the information I received from support, the issue should be resolved after updating QNAP AI Core to version 3.8.1.

I haven’t added many new photos since the update so I tested it by removing June’s photos from the source folder and then adding them back. Face recognition now seems to be working correctly.

I still need to repeat this remove-and-readd process for the rest of the affected photos.

9

I had already installed version 3.8.1 when I wrote the post.
The problem still existed at that point.

However, for some inexplicable reason, my QNAP started running “Detect Similar Photos” through the Multimedia Console again on its own this afternoon.
It only has 120,000 photos left to process.
I’m curious to see if things improve.

Update:

The version of the Multimedia Console was updated from 2.11.1 to 2.11.2.
The update apparently came this afternoon.
That’s why it’s starting the new AI scan.

10

Please post in the correct language

11

I apologize!
I didn’t mean any harm!
It seems my browser had automatic translation enabled, which made me think the thread was written in German.

Thank you for adding the English translation!