I use Digikam for face recognition on my large photo collection. Unfortunately, I’ve noticed that the face detection often makes mistakes, even recognizing images as faces when it’s clear to any human observer that they are not.
With version 8.5.0, which I’m now using on Linux, I was hoping for some improvements. However, I still have to spend a lot of time correcting errors manually.
I’m now looking for a way to completely rebuild the training model for face recognition. My idea is to delete the current model and generate a new one based on the data I already have: over 10,000 images with marked face regions and associated person tags. Ideally, I would like to run this training process overnight and start fresh with an improved model the next day.
Does anyone know how I can achieve this? Is there perhaps already a function for this that I may have overlooked, or some workaround to retrain the model from scratch?
If you look in the digiKam settings – “configure digiKam” – in the section “Database” there’s a setting for the location where the database files are stored.
digiKam stores data (including albums, album roots, tags, thumbnails, face recognition data, image metadata, file paths, settings and others) in four databases:
Face database for storing face recognition metadata: hosts face histograms for face recognition.
In that location there are four database files:
> ls /var/local/digiKam/xxx/
digikam4.db recognition.db similarity.db thumbnails-digikam.db
>
Quit digiKam.
Remove the ‘recognition.db’ file.
Restart digiKam.
I personally find that, the text in the configuration settings is better than the text in the online help pages quoted above.
Tx for your message. Removeing the recognition.db was not helpful. It removes me all already know faces as well and I have to go through all 100.000 pictures again. Goal is to use already named faces as new data for the training.
AFAIK, that’s what digiKam does anyway – albeit with a not so very strong algorithm as the competition and, AFAIK, without any support from an algorithm which uses a Large Language Model (LLM) …
Maybe in a future version the algorithm used by the Face Recognition feature will be improved.
