[digiKam-users] --> Tunning Face Recgnition

[Alex Antão]

Is there a way to tune face recognition better ? 

That's because it is finding some faces completely different from some people. Even if I apply them to the right one, new ones keep going to the wrong.

It also seems that it cannot forget some deleted faces. For example, I mistakenly applied a name to a face, that name was wrong and fixed it to the right. When I start the face recognition process, that deleted face apears again and the faces that should be assigned to the Corrected face, is assigned to the deleted one....

Thanks...

[woenx]

I remember there is a slider to adjust the sensitivity/specificity (more false positives/more false negatives) in the detection. But that won't improve the recognition by much (settings between 60 and 80 worked best for me). The fact is that the recognition only works well under certain conditions (a group of well defined faces looking to the camera, and without too many people to choose from). It seems that it will be improved during this summer as one of the projects for the Google Summer of Code 2020, so if it goes well, we may see the results by the end of the year.

Oh, and it's a know issue. If you delete a face rectangle, and run the detection again, it will be detected once again. There's a feature under development (again, summer of code) which will add an "Ignored" category that would solve this.

So I am optimist that in the future both issues will improve.

El 25/6/20 a les 21:29, Alex Antão ha escrit:

Is there a way to tune face recognition better ? 

That's because it is finding some faces completely different from some people. Even if I apply them to the right one, new ones keep going to the wrong.

It also seems that it cannot forget some deleted faces. For example, I mistakenly applied a name to a face, that name was wrong and fixed it to the right. When I start the face recognition process, that deleted face apears again and the faces that should be assigned to the Corrected face, is assigned to the deleted one....

Thanks...

[Maik Qualmann]

It is very important that you rebuild the training data for the new face
engine in digiKam-7.0.0.
Tools-> Maintenance...-> [x] Detect an recognize Faces-> [x] Clear an rebuild
all training data.
With this maintenance run, the face vectors are regenerated based on the
already confirmed faces.

Maik

Am Donnerstag, 25. Juni 2020, 21:29:54 CEST schrieb Alex Antão:

> Is there a way to tune face recognition better ?
>
>
>
> That's because it is finding some faces completely different from some
> people. Even if I apply them to the right one, new ones keep going to the
> wrong.
>
>
>
> It also seems that it cannot forget some deleted faces. For example, I
> mistakenly applied a name to a face, that name was wrong and fixed it to
> the right. When I start the face recognition process, that deleted face
> apears again and the faces that should be assigned to the Corrected face,
> is assigned to the deleted one....
>
>
>
> Thanks...

[Jens Benecke-2]

Hi,

if I choose to do this, will I have to reselect and reassign all faces?

I hope not. :-)

Jens

Am 25.06.20 um 21:49 schrieb Maik Qualmann:

> It is very important that you rebuild the training data for the new face
> engine in digiKam-7.0.0.
> Tools-> Maintenance...-> [x] Detect an recognize Faces-> [x] Clear an rebuild
> all training data.
> With this maintenance run, the face vectors are regenerated based on the
> already confirmed faces.
>
> Maik
>
> Am Donnerstag, 25. Juni 2020, 21:29:54 CEST schrieb Alex Antão:
>> Is there a way to tune face recognition better ?
>>
>>
>>
>> That's because it is finding some faces completely different from some
>> people. Even if I apply them to the right one, new ones keep going to the
>> wrong.
>>
>>
>>
>> It also seems that it cannot forget some deleted faces. For example, I
>> mistakenly applied a name to a face, that name was wrong and fixed it to
>> the right. When I start the face recognition process, that deleted face
>> apears again and the faces that should be assigned to the Corrected face,
>> is assigned to the deleted one....
>>
>>
>>
>> Thanks...
>
>
>

--
Regards, Jens

[fch22]

Hi Maik

I've done as you have indicated, unfortunaly after this maintenance run
, the face recognized is all except the expected ones (I'm using 7.0 RC)

by the way , I see that new faces are identified by a green frame but
didn't find how to filter them (to access directly to the non confirmed
faces)


thanks


frederic


Le 25/06/2020 à 21:49, Maik Qualmann a écrit :

> It is very important that you rebuild the training data for the new face
> engine in digiKam-7.0.0.
> Tools-> Maintenance...-> [x] Detect an recognize Faces-> [x] Clear an rebuild
> all training data.
> With this maintenance run, the face vectors are regenerated based on the
> already confirmed faces.
>
> Maik
>
> Am Donnerstag, 25. Juni 2020, 21:29:54 CEST schrieb Alex Antão:
>> Is there a way to tune face recognition better ?
>>
>>
>>
>> That's because it is finding some faces completely different from some
>> people. Even if I apply them to the right one, new ones keep going to the
>> wrong.
>>
>>
>>
>> It also seems that it cannot forget some deleted faces. For example, I
>> mistakenly applied a name to a face, that name was wrong and fixed it to
>> the right. When I start the face recognition process, that deleted face
>> apears again and the faces that should be assigned to the Corrected face,
>> is assigned to the deleted one....
>>
>>
>>
>> Thanks...
>
>
>

[Rob D]

I have a question about retraining the faces database.

Does the retraining look at all tagged faces including manually added faces, or just faces that have been detected and confirmed by the Digikam face recognition routines?  

I'm asking because I manually tag faces that really are not good to be used for training (obscured, bad angles, etc) and it would be counter productive to pollute the training data with this garbage data.

If all faces are used in training, maybe flagging manually added face regions would be a good idea (or flagging auto-detected regions so it works with old data coming forward).

Rob

On Thu, 2020-06-25 at 21:49 +0200, Maik Qualmann wrote:

It is very important that you rebuild the training data for the new face 

engine in digiKam-7.0.0.

Tools-> Maintenance...-> [x] Detect an recognize Faces-> [x] Clear an rebuild 

all training data. 

With this maintenance run, the face vectors are regenerated based on the 

already confirmed faces.

Maik

Am Donnerstag, 25. Juni 2020, 21:29:54 CEST schrieb Alex Antão:

Is there a way to tune face recognition better ? 

That's because it is finding some faces completely different from some

people. Even if I apply them to the right one, new ones keep going to the

wrong.

It also seems that it cannot forget some deleted faces. For example, I

mistakenly applied a name to a face, that name was wrong and fixed it to

the right. When I start the face recognition process, that deleted face

apears again and the faces that should be assigned to the Corrected face,

is assigned to the deleted one....

Thanks...

[Maik Qualmann]

All confirmed faces are used, including the manual ones. However, only the
last 100 are saved. If new ones are added, the older ones are removed. We
cannot store them all, and the Faceengine would slow down if it had to compare
thousands of face vectors. So this means that the quality of facial
recognition can currently vary. A student is working on improving facial
recognition as part of GSoC 2020. We expect further improvements in late
autumn.

Maik

Am Donnerstag, 9. Juli 2020, 16:15:52 CEST schrieb Rob Dueckman:

> I have a question about retraining the faces database.
>
> Does the retraining look at all tagged faces including manually added
> faces, or just faces that have been detected and confirmed by the
> Digikam face recognition routines?
>
> I'm asking because I manually tag faces that really are not good to be
> used for training (obscured, bad angles, etc) and it would be counter
> productive to pollute the training data with this garbage data.
>
> If all faces are used in training, maybe flagging manually added face
> regions would be a good idea (or flagging auto-detected regions so it
> works with old data coming forward).
>
> Rob
>
> On Thu, 2020-06-25 at 21:49 +0200, Maik Qualmann wrote:
> > It is very important that you rebuild the training data for the new face
> > engine in digiKam-7.0.0.
> > Tools-> Maintenance...-> [x] Detect an recognize Faces-> [x] Clear an
> > rebuild all training data.
> > With this maintenance run, the face vectors are regenerated based on the
> > already confirmed faces.
> >
> > Maik
> >
> > Am Donnerstag, 25. Juni 2020, 21:29:54 CEST schrieb Alex Antão:
> > > Is there a way to tune face recognition better ?
> > >
> > >
> > >
> > > That's because it is finding some faces completely different from some
> > > people. Even if I apply them to the right one, new ones keep going to
> > > the
> > > wrong.
> > >
> > >
> > >
> > > It also seems that it cannot forget some deleted faces. For example, I
> > > mistakenly applied a name to a face, that name was wrong and fixed it to
> > > the right. When I start the face recognition process, that deleted face
> > > apears again and the faces that should be assigned to the Corrected
> > > face,
> > > is assigned to the deleted one....
> > >
> > >
> > >
> > > Thanks...

[fch22]

Hi all


just a little up on my 2 questions:

- why the result of scan is showing all except the desired ones (after
rebuilding learning data from maintenance and running new scan with
option "recognized faces" and an accuracy at 92%)

- how to filter the new faces


thanks

frederic



Le 02/07/2020 à 19:31, frederic chaume a écrit :

> Hi Maik
>
> I've done as you have indicated, unfortunaly after this maintenance
> run , the face recognized is all except the expected ones (I'm using
> 7.0 RC)
>
> by the way , I see that new faces are identified by a green frame but
> didn't find how to filter them (to access directly to the non
> confirmed faces)
>
>
> thanks
>
>
> frederic
>
>
> Le 25/06/2020 à 21:49, Maik Qualmann a écrit :
>> It is very important that you rebuild the training data for the new face
>> engine in digiKam-7.0.0.
>> Tools-> Maintenance...-> [x] Detect an recognize Faces-> [x] Clear an
>> rebuild
>> all training data.
>> With this maintenance run, the face vectors are regenerated based on the
>> already confirmed faces.
>>
>> Maik
>>
>> Am Donnerstag, 25. Juni 2020, 21:29:54 CEST schrieb Alex Antão:
>>> Is there a way to tune face recognition better ?
>>>
>>>
>>>
>>> That's because it is finding some faces completely different from some
>>> people. Even if I apply them to the right one, new ones keep going
>>> to the
>>> wrong.
>>>
>>>
>>>
>>> It also seems that it cannot forget some deleted faces. For example, I
>>> mistakenly applied a name to a face, that name was wrong and fixed
>>> it to
>>> the right. When I start the face recognition process, that deleted face
>>> apears again and the faces that should be assigned to the Corrected
>>> face,
>>> is assigned to the deleted one....
>>>
>>>
>>>
>>> Thanks...
>>
>>
>>