how to optimally use face*tion

[Boudewijn]

Hi List,

TL;DR: how do I make optimal use of face detection/recognition? I seek to
improve recognition results.

I use DK, but not as a professional. Mostly family images. I am happy with the
face detection because it helps to identify and tag pictures with friends and
family on them quick and easy.

I think I do not use the most optimal way, so I am hoping for feedback.

Detecting faces, there are 4 kinds of false hits, and one that I want to tag:
1. detected as face, but it is a plant
2. detected as a face, but it is a painting of a face
3. detected as a face, but someone in the crowd I don't know
4. detected as a face, but I am not interested in tagging this specific
instance of the face of this acquaintance
5. detected as a face that I want to tag

One last category, I'm not sure where to put it: there is a face in the middle
of the thumbnail, but it is much less zoomed into than usual and got other
things around it (faces, plants, etc).

How do I help/disturb the training data most for
1. detection
2. recognition

1. I would guess that detection is helped by pruning the non-face false hits,
and keeping faces I am not interested in. I use the upper-right X on the face
thumbnail to reject it, and the lower-right 'no parking' symbol (circle with
diagonal slash) to make the thumbnail disappear from the unknown detections.

2. I would guess that recognition is helped by having a small number of
disperate face tags to choose between, and a large number of confirmed faces
with each name.

Lacking time to hang as many tags on each image as I would like, I kept, in
the end, to a minimum of tagging people (the old fashioned way) and places
(hardly geo-tagged) as a minimum for making fast intersections between
combinations of people and visited places (fully realizing those would be the
first to be automated and making it futile to spend time on those...)

So far face detection helps me, but face recognition does not yield the
results I hoped for so I end up going through the set of images more than once
to check for false negative names, false positive names and just for fun.

Maybe I should mention that making the process do anything at all is no
problem at all after selecting some albums or tags, moving the parameters left
or right and selecting or deselecting the "All cpu's are belong to DK - all
cpu's - all cpu's are belong - .. sorry, couldn't leave it ;-) ).

Thanks for reading through all of this. Thanks even more for everyone involved
in bringing digiKam!

Best regards,

Boudewijn

PS: should the parameter moved to "high accuracy", like 95%, (only) make fewer
false positive detections, or does it (also) increase false negatives? And
does it make training data into garbage if I run it in low accuracy once? In
other words: fast will give me (only) results that are faces for sure after a
quick scan (skipping difficult faces and fetching plants), and accurate will
give me "hardly any" plants and "will not" skip difficult faces?
Will this improve after training, or is each image scrutinized on its own?

[AndriusWild]

In your case I would started with the default accuracy settings (1/2) and if keep getting plants and unwanted faces would moved the slider to 1/4 and reviewed the results

Sent from my Samsung Galaxy smartphone.

-------- Original message --------

From: Boudewijn <[hidden email]>

Date: 2017-02-28 3:48 PM (GMT-07:00)

To: digiKam - Home Manage your photographs as a professional with the power of open source <[hidden email]>

Subject: how to optimally use face*tion

Hi List,

TL;DR: how do I make optimal use of face detection/recognition? I seek to
improve recognition results.

I use DK, but not as a professional. Mostly family images. I am happy with the
face detection because it helps to identify and tag pictures with friends and
family on them quick and easy.

I think I do not use the most optimal way, so I am hoping for feedback.

Detecting faces, there are 4 kinds of false hits, and one that I want to tag:
1. detected as face, but it is a plant
2. detected as a face, but it is a painting of a face
3. detected as a face, but someone in the crowd I don't know
4. detected as a face, but I am not interested in tagging this specific
instance of the face of this acquaintance
5. detected as a face that I want to tag

One last category, I'm not sure where to put it: there is a face in the middle
of the thumbnail, but it is much less zoomed into than usual and got other
things around it (faces, plants, etc).

How do I help/disturb the training data most for
1. detection
2. recognition

1. I would guess that detection is helped by pruning the non-face false hits,
and keeping faces I am not interested in. I use the upper-right X on the face
thumbnail to reject it, and the lower-right 'no parking' symbol (circle with
diagonal slash) to make the thumbnail disappear from the unknown detections.

2. I would guess that recognition is helped by having a small number of
disperate face tags to choose between, and a large number of confirmed faces
with each name.

Lacking time to hang as many tags on each image as I would like, I kept, in
the end, to a minimum of tagging people (the old fashioned way) and places
(hardly geo-tagged) as a minimum for making fast intersections between
combinations of people and visited places (fully realizing those would be the
first to be automated and making it futile to spend time on those...)

So far face detection helps me, but face recognition does not yield the
results I hoped for so I end up going through the set of images more than once
to check for false negative names, false positive names and just for fun.

Maybe I should mention that making the process do anything at all is no
problem at all after selecting some albums or tags, moving the parameters left
or right and selecting or deselecting the "All cpu's are belong to DK - all
cpu's - all cpu's are belong - .. sorry, couldn't leave it ;-) ).

Thanks for reading through all of this. Thanks even more for everyone involved
in bringing digiKam!

Best regards,

Boudewijn

PS: should the parameter moved to "high accuracy", like 95%, (only) make fewer
false positive detections, or does it (also) increase false negatives? And
does it make training data into garbage if I run it in low accuracy once? In
other words: fast will give me (only) results that are faces for sure after a
quick scan (skipping difficult faces and fetching plants), and accurate will
give me "hardly any" plants and "will not" skip difficult faces?
Will this improve after training, or is each image scrutinized on its own?

[Boudewijn]

On dinsdag 28 februari 2017 16:10:40 CET Andrey Goreev wrote:
> Boudewijn wrote:
> > TL;DR: how do I make optimal use of face detection/recognition? I seek to
> > improve recognition results.

> In your case I would started with the default accuracy settings (1/2) and if
> keep getting plants and unwanted faces would moved the slider to 1/4 and
> reviewed the results

Is that applicable to detection or recognition? At 95% about 1 in 15 is a
"clear" false positive (ie, plant, not a painting with a face on it); I never
checked for false negatives: I got many faces as it is.

> > So far face detection helps me, but face recognition does not yield the
> > results I hoped for so I end up going through the set of images more than
> > once to check for false negative names, false positive names and just for
> > fun.

Can I expect the recognition rate to increase with a growing number of already
known faces?

Thanks and best regards,

Boudewijn