0.6.2 — 2026-07-25
- History scan can recover missed events (
python -m app.backfill --rescue). When the
detect snapshot holds no usable face, the event clip is scanned instead — measured over
180 events, about one in five yields a face that the normal scan misses entirely. - Quality gate calibrated against real data, not intuition. A first run put 308 faces
into the review queue, most of them useless: back-of-head shots, motion blur, and
outright false positives (a church spire scored 0.57). Sorting 74 finds by detection
score showed a clean split — below ~0.8 it is mostly junk, above it mostly real faces.
Rescue therefore requires 0.85 by default (--rescue-min-det), which cuts the yield
to roughly a seventh and leaves the usable finds. - Deliberately no filter on gallery match score: a stranger's face scores low by
definition, and enrolling strangers is what the queue is for. Also no pose filter —
measurement showed working galleries are full of profile shots (median frontality 0.55),
so filtering those would discard good references.