Perhaps off-topic but long ago I read that faces are better detected in the YCgCr color space:
Hi @cedric,
Thanks for the link. It’s an interesting article. Facial recognition algorithms have advanced significantly in the 13 years since the article was published. The foundational algorithm used in the paper has long been surpassed by more efficient and accurate models. Modern facial detection and recognition models are based on machine learning models, and it’s important to use the same color space as the images used to train the model. For most modern models that’s either RGB or BGR.
Cheers,
Mike
3 Likes
This is a bit of an older topic, but I’ll revive it to say a huge THANK YOU to @Michmill! DigiKam’s face recognition is now operating on a whole new level!
Where in the past it would recognize perhaps one in three faces correctly, the hit rate was closer to 80% in my latest batch. That’s fantastic stuff! Thank you so much!
4 Likes
Hi @bastibe,
I’m glad it’s working better for you. What version of digiKam are you using?
I’m happy that it’s better, but I’m a little concerned you’re only get 80% accuracy. You should be seeing better than 90%, and hopefully closer to 95%.
Cheers,
Mike
1 Like
Now that you mention it, I was using the previous version, v8.5.
The 80% figure was just estimated very roughly, mostly because it just didn’t find some faces. But what faces it found were mostly attributed correctly. And where it couldn’t attribute them correctly, it left them blank, which is a huge improvement over the many mis-identifications I had before.
But I’ll download the newest version right now and have a go with that!
(Side note, a long while ago, the face recognition in Digikam annoyed me enough to look for commercial alternatives. I could not, however, find a better one. What we have now in Digikam is better than what I saw in e.g. Lightroom, ACDSee, or Mylio)
Picassa was awesome at facial recognition.
I agree. Picassa was great. I’m using Picassa as my benchmark when it comes to improving facial recognition in digiKam. We’ve add more features in the upcoming 8.7.0 release of digiKam to make the experience more like Picassa. Features like background facial recognition and saving rejected suggestions so the next best match will be presented.
Cheers,
Mike
4 Likes