Is it worth trying to sort out 4k+ face clusters?

[AstroPhotoJay]

I've been tinkering with Nextcloud and Recognize for a while now, with a focus on using Recognize for facial recognition, and have reached a point where face clustering has turned into a disorganised mess and I'm not sure if it's worth any more effort.

In NC, I have
274,246 indexed media files in Memories
336,742 face detections
4,117 face clusters
81,482 unassigned faces

Of these 4k face clusters, there are approx 40 which contain only one image. The largest unnamed/unchanged clusters only contain about 100 images. Most of the 4k clusters have somewhere between 10 - 70 faces.

There are countless duplicate clusters for the same people, images of myself and my partner (who looks nothing like me) have been grouped into clusters together and countless shitclusters.

To be fair, the content library is messy. It spans 20+ years and the images would be challenging to work with (people getting older, images of people with and without sunglasses on, at costume parties, partially covered faces, etc...)

I don't want to dump recognize just yet, as I have spent more time than I care to admit manually adding/removing/merging clusters in the hope that accuracy would improve. But at the same time, I don't see how its feasible to sort through this mess.

Does any one have recommendations on how I might go about improving the accuracy of my clusters and sorting through this mess without throwing out all of the manual effort? Would this even be worth doing?

Any tips/tricks much appreciated.

[marcelklehr]

Hey @AstroPhotoJay

I'm sorry you're disappointed by the quality of the face recognition. Manually assigning and merging clusters is only going to get you so far, currently, because the model that extracts the face vectors leaves a lot to be desired. If the automatic run is already generating too much garbage, I don't think it is worthwhile to correct it. We are currently planning a rewrite of this app in python in order to be able to use newer models that will greatly improve the quality of the clusters. I recommend waiting until this effort bears fruit :)

Cheers
Marcel

[AstroPhotoJay]

Thanks for taking the time to reply - it's very much appreciated. Since I have already done some work manually assigning and merging clusters, I have about 40 named faces containing anywhere from a few hundred photos to the largest having 9.8k. Do you think I would get better results if I were to delete all un-named faces from the oc_recognize_face_clusters table and then run occ recognize:cluster-faces?

[marcelklehr]

I don't recommend messing with the database manually. I also doubt this will drastically improve results.