Could you not use metadata to create a confidence score? For example, if two people look similar but the dates of the photos are 100+ years apart, it’s unlikely they’re the same person.
This is how we find the time travelers :)
This is how we find the time travelers :)
Date the photo was taken.
Where the photo was taken (e.g. the annual family reunion in SmallTown, Illinois)
Photographer. They're likely taking photos of the same general set of people or in the same geographic area.
Time+location of other similar photos. Back in the days of the civil war, it wouldn't be possible for someone to be in to locations 1,000 miles apart in 2 days.
Side note: I recently saw a video from ~1910 in New York City. Everybody in the video had a similar physical build and would make the "average" person today look fat. I wonder how this would affect the confidence of photo grouping? Overall, did people look more homogeneous back then?
If your goal is explicitly to find something extremely rare - unexpected face matches in different places - you will be flooded with false positives even if the false positive rate is fairly low.