https://bugs.kde.org/show_bug.cgi?id=291514 Summary: Suggestions for improving face recognition
performance.
Product: digikam
Version: unspecified
Platform: Unlisted Binaries
OS/Version: Linux
Status: UNCONFIRMED
Severity: normal
Priority: NOR
Component: Face Management
AssignedTo:
[hidden email]
ReportedBy:
[hidden email]
Version: unspecified (using Devel)
OS: Linux
Digikam's face recognition performance is reported to be well below Picasa's
implementation.
Since this is a 2 part problem (face detection + face recognition) I would like
to focus on the second one: Recognition.
This is typically a Supervised Machine Learning problem than can benefit a lot
from as many information as it can gather or has already gathered before. One
source is the image itself (and I've read that it uses OpenCV to extract face
characteristics so in this field I have nothing to comment).
What I would like to suggest is (in case it's not done already) that:
* People already identified on a training set should have their probability
boosted for recognition if the date of the test set is very close (I mean we
know they were there in the shoot)
* People whose photos are in the same folder as the training set (already
identified ones) should have their probability boosted for recognition (because
we already manually organize them that way)
* People who tend to cluster in the training set should have their probability
boosted for recognition (if there is an unidentified person on a picture where
I already identified my father there is a big chance that it is my mother)
* People whose photos are geographically clustered should have their
probability boosted for recognition if the photo is taken in a close location
(if the picture was taken in my hometown my parents are mre likely to be in it)
Reproducible: Didn't try
Steps to Reproduce:
NA
Actual Results:
NA
Expected Results:
NA
NA
--
Configure bugmail:
https://bugs.kde.org/userprefs.cgi?tab=email------- You are receiving this mail because: -------
You are the assignee for the bug.
_______________________________________________
Digikam-devel mailing list
[hidden email]
https://mail.kde.org/mailman/listinfo/digikam-devel
Locked
