Sorting Face Recognition by Recognition Score
david.s.rothman at gmail.com
david.s.rothman at gmail.com
Wed Dec 13 15:34:28 GMT 2023
I'm working through 250k+ pics that I'm trying to process for a family member.
The main problem is that there is a huge percentage of dupes scattered thru the picture store (even after pre- addressing b4 using Digikam). The secondary problem is a huge number of kid/baby photos.
With Digikam, I trained the algo on a small number of folders of adults only and then chose a very high specificity number. Over the past week, I've relaxed that number and somewhat randomly added addl matches. Then I'd go back to a more sensitive number. Rinse and Repeat. I've now done this probably 20 times and it feels like I'll need to do it another 50x before I get to the point where we capture most of what we want.
It is safe to assume that rejecting a suggestion doesn't do anything other than rejecting that suggestion at this point in time. IOW, there is no lasting impact on that action so that that pic can and will show up on future searches?
Two suggestions:
1. The Algo must provide a score (or perhaps a set of scores) when it goes thru face recognition. When the results are presented to the user, it would be very helpful to allow us to sort by that score or scores. That way when 1000+ results are displayed, we don't need to hunt for the good/obvious matches (or similarly, reject the blurry unrecognizable ones). That would be an incredible time save to users who are trying to process a lot of data.
2. When processing a lot of the face recognition results, a slip of the mouse can and does mark pics incorrectly - in some cares a selection of many pics. Correcting via ^Z doesn't seem to work (or if it does, it isn't obvious to me). I'm especially concerned about the false positives. I don’t know if the Digikam saves date/time for each recorded recognition, but if so, sorting on that would allow the user to correct mistakes much more easily.
Thanks for a great program and any comments/help.
More information about the Digikam-users
mailing list