Face recognition outside IMatch, offline

Started by mastodon, January 07, 2024, 09:52:43 PM

Previous topic - Next topic


I have a lot (more than a thousend) photos, that are older than 50 years, even 100 years from my family history. Lot of different people, who are mostly relatives. As I hardly find anybody in the family, who can recognise this people, I want to use a face recogntiion software. The siblings, brothers and mothers-sisters etc very hard to recognise. A have a terrible headache to recognise them, even on group photos.
As IMatch is not very good at FR yet (especially not front faces), do you use any oder FR seervices? There is Picasa, that is very good. A have read that digiKam is good at that. And maybe someone is using CompreFace or Deepstack even with collaboration (with the use of Double Take).
Does anybody use offline FR software in windows? What are the experiences?


QuoteAs IMatch is not very good at FR yet (especially not front faces),
IMatch is actually considered to be quite good at face recognition. In the standard benchmarks it comes out at 96% and better, the last time I've looked... Are your sure your trained faces in IMatch are good?

If there is only one eye in an image, the face is partially obscured or only a small part of the face is visible, all AIs will 
have a hard time. Unless they are very specialized and trained for exactly that purpose. But then they will fail on more normal faces...

Have you actually tried the applications and services you mention? Do they do a good job for your images?
In my experience, if even humans struggle to recognize faces (after 200,000+ years of training), AIs struggle even more.
Give the systems you mention a try and if they work for you, awesome. Let us know what you find out.

As long as the software or service you use can persists its results in standard XMP face regions, IMatch will import the info and make it usable for you like the faces IMatch has detected itself.
-- Mario
IMatch Developer
Forum Administrator
http://www.photools.com  -  Contact & Support - Follow me on 𝕏 - Like photools.com on Facebook


I have cleaned up the trained faces pool, and that made recognition better, although there are some strange cases, when IMATCH does not even recognise a face as face, that is quiete "clear". BUT, on clean, front faces IMATCH work very good.
I have tried only Picasa, that is why I asked for the others.


I have installed Picasa, and well, IMatch is about the same in recognising faces, but Picasa is a bit better in detecting faces (no false positives, with treshold: 80-80).
I think the face grouping by similarities, that helps make recognition easiar. But it does not worth to use with IMatch. So, IMatch got defenitely better by the years in this matter :)
I have installed digitKam 8.2, but could not run, crashed many times. This is a more than 100 years old photo, not very good. The face in the middle was detected by Picasa, but failed by IMatch.


Testing. So, IMatch is better in face recognition, than Picasa. I think, because it uses time data. Clever.


Picasa was bought and then abandoned by Google many years ago - in favor of their cloud-based replacements.
Google favors solutions where they can get your data into their cloud, to analyze it using their AIs and then monetize the sh*t out of it. Software running offline does not generate money for them - so it's of no interest to Google.

QuoteThe face in the middle was detected by Picasa, but failed by IMatch (...)
Picasa may be better in some circumstances. IMatch may be in others.
You don't show your reference image so I cannot say more about this.

Recognizing faces out-of-focus, partially obscured faces or even people facing away from the camera is, in theory, possible. You'll have to train AI models specifically for that purpose, using hundreds of thousands of input images and spending weeks or months of very expensive cloud computing on it.
And even then you'll might end up with a 40% to 70% recognition rate. And probably a much worse recognition rate for persons facing the camera...

Us humans have been trained to tell friend from foe and friends & family from strangers for ~200,000 years. Our mushy brains are actually very, very good at tasks like face recognition or sentiment analysis. AIs are not nearly as good yet.

The face AI I've implemented for IMatch is very good. And I've came up with some additions, like considering the age of persons or allowing users to pick the images to train the AI with.
This improves things, but I think there will never be a 100% recognition rate.
Even if you upload all your images to the Google, Adobe or Microsoft clouds and pay money for each batch of images, they're only right in 60 to 90% of all cases. Which gives you between 1 out of 10 and 6 out of 10 false results.
-- Mario
IMatch Developer
Forum Administrator
http://www.photools.com  -  Contact & Support - Follow me on 𝕏 - Like photools.com on Facebook