You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
New: references that do not resemble their own person are set aside automatically.
Version 0.14 checked every reference against the other people in the gallery. That
check has a blind spot, and it cost a misidentification here five days later.
What happened: one photo in my own gallery — taken from above, showing mostly the back of
a head — was published under my name when a different person walked in. It reached 0.576 against her live face, well past the threshold of 0.45.
The cross-check had seen that photo and let it pass at 0.368, just under its limit.
It was not wrong; it measures gallery against gallery, and a live face comes far closer
than any stored photo of the same person does. Gallery-vs-gallery is a lower bound, and
this case shows by how much.
The stronger signal was there all along: that photo's average similarity to my own other
49 photos was −0.011, where 0.265 is normal for me. It was the single
worst-fitting photo in my own gallery, and needed no second person to be recognised as
worthless. A reference that barely resembles the person it belongs to holds almost no
face information — and an embedding without face information matches strangers as
readily as its owner.
FaceID now applies both checks in one pass: on new photos, once at startup, and on demand
(Settings → Keeping the gallery clean → check gallery now). Verified against the
real event: with the photo set aside, that same face no longer matches me at all — I drop
out of the top three — and it is correctly recognised as the right person at 0.460.
Measured relative to each person, not by a fixed number: how alike someone's photos
normally are varies a lot with how varied the shots are (0.23 to 0.47 across the people
here). A photo falls when it stays below self_outlier_ratio (default 0.25) of that
person's median, and only for people with at least 5 photos, where the median means
something. Across 188 photos of 13 people it set aside 3 — the culprit above, one
blown-out IR frame, one more head seen from above.
self_outlier_ratio: -1 switches it off. As always nothing is deleted: photos move to
the person's set-aside pile marked ⚠ with the numbers behind the decision, and one click
restores them.