librephotos
librephotos copied to clipboard
LibrePhotos
Face recognition feedback from a new user perspective
Describe the enhancement you'd like This may not fit the usual enhancement request. I'd like to share my experiences with the face recognition workflow.
Describe why this will benefit the LibrePhotos Feedback from a user perspective may help development.
Additional context I'm amazed by the accuracy of the face recognition. It often even detects people in blurred backgrounds. This also has the following consequences to which I'm not yet certain how to deal with:
Sometimes the algorithm detects faces from art (paintings, statues, etc.). There's an option to move to 'Other/Unknown' but as they are not actual people I'm often deleting the face and the thumbnail. I'm not sure which kind of algorithm/training is used, would it be beneficial to train the algorithm that a detected "face" is incorrect instead of just deleting it (or does this already happen)?
About the same goes for random unknown people being detected in the background. Of course there's 'Move back to Other/Unknown' but I think it would be nice if you could differentiate between 'Hey, this is someone else' and 'This is someone else, but I do not care who it is'. This way if person B is inferred to be person A, I can move person B back to Other/Unknown and tag person B later (kind of like a backlog). But if person C would be inferred to be person A and I don't know this person, I could remove the association and possibly train the algorithm.
PS: It seems like the 'Label faces' modal dialog has bugged person pictures. The wrong faces show up with the listed persons in the dialog that shows when clicking the green plus symbol. I haven't been able to look into why though. It seems that they shuffle every now and then.
would it be beneficial to train the algorithm that a detected "face" is incorrect instead of just deleting it
Yeah this sounds nice, I have such cases too. I would imagine it could work like a negative prompt for generative NNs, but I'm not sure if it's possible from a technical standpoint.
