face recognition in digiKam 7.0.0-beta1

Hi there,

I’m testing face recognition new implementation on Beta1.
I have no positive recognition on the non-confirmed tag even with teaching the algorithm up to 10 images.

does it work for someone?
I did a detection of face, then tag some pictures and then run the recognize face feature. but no result appears.
working on a small db for test, it took some how 2 minutes of calculations.

1st of all is my workflow correct?

Yes, that is the way to do it and it works a lot better than it used to.
Things like the size of the face in the image and lens distortion do matter.

I’m also seeing behavior I don’t quite understand in regards to the “Uncomfirned” tag. Sometimes no faces turn up after a run. I’ve not had time to look into it properly and my situation is complicated by>

  1. Having used the previous face detection/recognition tool on some of the same images so there might be old cruft in the database preventing scan? Not all were recognized or detected with the previous algo.
  2. Running the tool by right clicking an album or doing the Workflow|Search in|Settings in the lower left of the People tag.
  3. Sometimes running with “detect faces” and sometimes with “recognize faces”.

Restarting digikam seems to (not sure yet) populate the “Unconfirmed” tag.

I haven’t found any documentation about how the settings are expected to influence results. I you need to run both detect and recognize to have auto recognition of faces.

The new tool is very good at detecting faces which is great!

Hello, are you running beta 1 (uploaded on Dec 22nd)?
If yes could you give a spin to a daily build (Jan 14)?

@Andrius es I’m running Beta 1 from 12/22

I am in the same case study as @nosle in my second trial. I used a copy of my production database (around 4000 pictures) and it was tagged with the “old system” (6.X)
So I have analyzed again the pictures and rebuilt the deep learning db, and it works.

So I think the original issue come from the fact that pictures where tagged or analyzed (I don’t know) in the previous way.

so looks like it works pretty well


1 Like

so following this topic and the use of face recognition, I’m facing an issue.
When I have a hierarchy such as “people/family/person1” and I run face recognition, some unknown people are suggested as “family”, and not as a person.
so if intermediate tag could be removed by defoult would be good.

I also see that the more you have face of a person, the less digikam can identify him.
a person with 10 photos before running face recognition is more proposed in “non confirmed” tag than one with already 2500 pictures

In the people left sidebar family should show up as a person. Then in the context menu there is an option to make this not a person anymore (don’t know by hart, should be easy to recognize).

I see the same thing - I suspect some model overfitting problem. I have an approach in mind that targets the workflow/UX, but don’t have the time to implement. Essentially it comes down to not only save the “best match”, but all matches above a threshold. Then the same unconfirmed face can turn up under several people. Once any of those is confirmed, all other matches would obviously get wiped.

Looks like there are multiple bug reports on face recognition:

Bugzilla search for face recognition

@Nemesis do you mind posting your comment in on the bug reports whichever you feel is more relevant to you?

And indeed there already is a report for exactly this problem: https://bugs.kde.org/show_bug.cgi?id=415782 :slight_smile:

A question/suggestion about the workflow. After running face recognition and the Unconfirmed tag is populated I can only see the faces in the Thumbnail layout. I can’t see what face digiKam has guessed/identified without clicking the thumbnail. This means I can’t quickly confirm digiKams guesses.

Is there an ideal workflow for facetagging that avoids this problem?

If not I have a few suggestions.

  1. Show the tag under each identified face in thumbnail view (Unconfirmed folder)
  2. Allow sorting faces by identified tag. By having faces grouped it’s much quicker mentally and physically to manage tagging
1 Like