Let me proper understand how recognize works...

[avDuma]

Hi all,

In these days I'm trying to use recognize face detection for my library of approx 50k pictures.

I'm running Nextcloud from TrueNAS docker, using 2 CPU and 6 memory of GB enabled for the whole container and 4 GB as memory limit.

Sometimes in the Nextcloud log I've receive some errors like the following ones:

[core] Fatale: Request used more than 300 MB of RAM: 1.3 GB - dal ? da -- alle 12 ago 2025, 07:51:03
[cron] Avviso: Cron job used more than 300 MB of ram after executing job OCA\Recognize\BackgroundJobs\ClassifyFacesJob (id: 173118, arguments: {"storageId":3,"rootId":206}): 1.3 GB (before: 33.1 MB)  - dal ? da -- alle 12 ago 2025, 07:50:40

First question:

From nextcloud docker is it possibile to increase the 300 MB thershold for cron job? Seems something already handled, but still not enough.

In order to better understand the whole process and thanks to the documentation provided, I've done a occ recognize:clear-background-jobs -> see on the App settings from Admin user:

Face recognition

Face recognition is working.
Riconoscimento delle facce: 0 Queued files, Last classification: 20 ore fa, Scheduled background jobs: 0, Last background job execution: 4 minuti fa
Face clustering: 0 faces left to cluster, Last clustering run: 19 ore fa, Scheduled background jobs: 0,
A minimum of 120 faces per user is necessary for clustering to kick in

I've thought that means that there's no process scheduled, so I've launched a occ recognize:recrawl

And the queued files on "face recognition" increase until the number 48492 files, than start decreasing -> maybe files are processed in order to recognize faces, but during face clustering, the "number of faces left to cluster" remains 0.
Strange is that no faces remains on face clustering: do you have an idea why?

Also, other strange behavior:

• There's some unassigned faces that belongs clearly to a cluster with 300 faces on it and still I have to assign manually to that person
• In some pictures detects some faces on the background, and not the main ones on the foreground, even if they are clear

Is it possible that my "face library" is too big for the server resources and when recognize "tries to open it on RAM" it exceed some memory limit and cannot cluster anymore?

In order to avoid other memory limit, I've reduced from 500 to 50 the file processed for run, even if recognize tell me that WASM mode isn't needed and I'm using standard TensorFlow.

If you can suggest me other direction for improve the whole process, it will be perfect for me, too. Thank you in advance!

[marcelklehr]

maybe files are processed

Yes, it seems that the files are being processed if the numbers decrease :)

Strange is that no faces remains on face clustering: do you have an idea why?

It could be that the faces in your photos are not detected for some reason.

Is it possible that my "face library" is too big for the server resources and when recognize "tries to open it on RAM" it exceed some memory limit and cannot cluster anymore?

This is unlikely. In the case of large clusters recognize draws a random sample from the cluster to compare it with new faces, but of course this could lead to a non-match, when it should be a match...