0.6.4 — 2026-07-25
- New:
scripts/coverage.py— per person, how well is she actually covered and what
is missing? Reports photo count, diversity (mean pairwise similarity), viewing angles
estimated from the landmarks (frontal / half / profile), day vs night/IR shots, and a
leave-one-out self test — then names the concrete gap ("no frontal shot", "no night
shot", "too few photos"). - New:
scripts/measure-recognition.py— did enrolling actually help? Compares the
current gallery against an older one (an unpacked backup) via leave-one-out, plus a
practical probe against recent Frigate events. - The self test is deliberately harsh and pessimistic on small galleries — with five
varied photos, each has to hold up against four entirely different situations. It is
only reported as a defect from eight photos up, and a low value means "these photos
reinforce each other weakly", not "this person is not recognised in practice".