Bug 415895 - DNN Face recognition not working well on a picture
Summary: DNN Face recognition not working well on a picture
Status: RESOLVED FIXED
Alias: None
Product: digikam
Classification: Applications
Component: Faces-Recognition (show other bugs)
Version: 7.0.0
Platform: Microsoft Windows Microsoft Windows
: NOR normal
Target Milestone: ---
Assignee: Digikam Developers
URL:
Keywords:
Depends on:
Blocks:
 
Reported: 2020-01-05 14:03 UTC by Alexandre Belz
Modified: 2020-09-03 21:01 UTC (History)
2 users (show)

See Also:
Latest Commit:
Version Fixed In: 7.2.0
Sentry Crash Report:


Attachments
Sample picture (61.51 KB, image/jpeg)
2020-01-05 14:03 UTC, Alexandre Belz
Details
Detection OK (179.73 KB, image/jpeg)
2020-01-05 14:04 UTC, Alexandre Belz
Details
Manual Identification of one face (140.97 KB, image/jpeg)
2020-01-05 14:04 UTC, Alexandre Belz
Details
Step 2 : manually identified face disappears from unkown. OK. (183.24 KB, image/jpeg)
2020-01-05 14:06 UTC, Alexandre Belz
Details
DNN does not recognise Megan on the duplicated picture. Test is KO. (175.78 KB, image/jpeg)
2020-01-05 14:07 UTC, Alexandre Belz
Details
DebugView Log file for DNN recognition. (23.10 KB, text/plain)
2020-01-05 14:08 UTC, Alexandre Belz
Details

Note You need to log in before you can comment on or make changes to this bug.
Description Alexandre Belz 2020-01-05 14:03:17 UTC
Created attachment 124903 [details]
Sample picture

SUMMARY
A face is not dectected on a duplicated picture while I identified it on the main picture.

STEPS TO REPRODUCE
1. Start from a cleaned DNN database 
Download the picture enclosed. Duplicate it. Scan for faces. See attached screen capture for normal detection results in people view.
2. Identify one face (Megan here). See attached screen capture.
3. Re-launch the recognition process (only to that directories)

OBSERVED RESULT
The other identical face of Megan is detected but NOT recognized (it remains empty/unkown).

EXPECTED RESULT
DNN should easily recognize the same face on the same duplicated picture !

SOFTWARE/OS VERSIONS
Windows: 10
digikam Build date: Jan 4 2020 (target: RelWithDebInfo)
Rev.: eef68a08113568f422583e9437dd2f459cb4bf79
Comment 1 Alexandre Belz 2020-01-05 14:04:08 UTC
Created attachment 124904 [details]
Detection OK
Comment 2 Alexandre Belz 2020-01-05 14:04:42 UTC
Created attachment 124905 [details]
Manual Identification of one face
Comment 3 Alexandre Belz 2020-01-05 14:06:24 UTC
Created attachment 124906 [details]
Step 2 : manually identified face disappears from unkown. OK.
Comment 4 Alexandre Belz 2020-01-05 14:07:35 UTC
Created attachment 124907 [details]
DNN does not recognise Megan on the duplicated picture. Test is KO.
Comment 5 Alexandre Belz 2020-01-05 14:08:15 UTC
Created attachment 124908 [details]
DebugView Log file for DNN recognition.
Comment 6 Maik Qualmann 2020-01-05 14:23:09 UTC
You don't understand DNN. After you have saved a "pattern", you expect that "pattern" will be found again immediately in the same image. This is how the old algorithm worked. DNN means learning, learning like children...
Give the engine more images of Megan and she will recognize the other image. I have 2 young girls here, they were often confused even in real life. DNN is always 99% correct for a wide variety of images.

Maik
Comment 7 Maik Qualmann 2020-01-05 14:26:38 UTC
You can also try reducing the initial accuracy 45-60.

Maik
Comment 8 Maik Qualmann 2020-01-09 17:48:50 UTC
Git commit 9d60f152e77b5201ce6bf23473684a32444bf586 by Maik Qualmann.
Committed on 09/01/2020 at 17:48.
Pushed by mqualmann into branch 'master'.

fix wrong item selection for face recognition
Related: bug 416018, bug 416028, bug 415782, bug 415867

M  +3    -1    NEWS
M  +0    -29   core/utilities/facemanagement/widgets/facescanwidget.cpp
M  +15   -0    core/utilities/maintenance/facesdetector.cpp

https://invent.kde.org/kde/digikam/commit/9d60f152e77b5201ce6bf23473684a32444bf586
Comment 9 Alexandre Belz 2020-01-09 17:57:29 UTC
Thank you Maik !
do you need me to check if it works better now ?
Comment 10 Maik Qualmann 2020-01-09 20:04:25 UTC
The bug was not yet closed with the commit, the message was also posted here so that users can test whether the problem has been resolved. But no packages containing the change will be available before tomorrow evening.

Maik