I’ve been ignoring this for a while, but it’s starting to get annoying, so…
Recognise, on the face of it, seems to be installed OK - all the things in the recognise dialogue are green / OK.
Trouble is, the manual seems to be one page long, documents nothing, and it doesn’t seem to really work.
I have a suspicion that it operates on the thumbnails of images - something I have limited to 128x128 because of the sheer waste of disk space that enabling all thumbnail sizes causes. I only ever want small thumbnails, or the full quality image.
But in regardless of suspicions, the results are awful, and it seems no longer to be recognising people in new images. Perhaps this is related to the message in recognises setting page that says:
"Face clustering: 92 faces left to cluster, Last clustering run: 20/08/2026, Scheduled background jobs: 0,
A minimum of 120 faces per user is necessary for clustering to kick in "
?
It seems to have been saying this for a long time now. But why? I don’t want to take 120 pictures at a time, just to have it classify one face. I’ve already got tens of thousands of images! surely it can try to classify one new image once in a while, without needing another 119 queued up for, I dunno, LOLs?
Whats the deal? If the manual had any useful content, maybe I could figure it out, but it doesn’t (AFAICT).
It’s very possible that by limiting the thumbnails to 128px you are also limiting recognize to 128px, as it uses the thumbnail generator to downsize the images in your nextcloud for classification.
As for the face clustering issue, it could be that the background job that clusters images runs out of memory (worth checking the logs for that), but it seems to have run on Aug 20. You can try doing a manual clustering run by doing `./occ recognize:cluster-face -b 10000` on the terminal of your server.
Ah, sounds like my suspicion on thumbnails was right… Is there any way to force recognize to, well, not do that, I guess? ie. just take the original images, unscaled, or take the originals and scale them to some known size without a requirement for generating gigabytes of previews I don’t want?
Re: clustering - I see no evidence of it running out of memory - the machine has 128GiB of ECC RAM and 30GiB of fairly fast swap.
Forcing a manual clustering run as you suggested, produced no errors, but several lines saying:
ClusterDebug: Not enough face detections found
The Web UI still says exactly the same as before - even the date has not changed. I assume it has just given up and done nothing, even though there are still 92 images to process…
Recognize has a fallback to use gdlib if the preview generator is not available, but by default it will use the preview generator, sorry.
If the manual clustering run didn’t change anything, you likely have a user with only 92 images. That is too few to do a proper clustering, hence it just leaves them be.
fyi: In nextcloud 35, recognize comes with an optional python backend as an exApp. That code works without the preview generator in the middle, so should produce much better results for you, if you’d like to try it.