SkyTechNerds/faceid v0.13.1

on GitHub

Three fixes from a detailed field report in #8 — all of them in the log rather than in
recognition itself, but a log that misleads is worse than a terse one.

• "largest face 54px < min_face_px 48" was a contradiction, and it was ours. The
  message only ever reported the width, while best_face() also requires the height and
  a minimum detection score. A 54x40px face failed on height, a 54x54px face with score
  0.40 failed on the score — both were reported as "too small". The line now names the
  actual reason: largest face 54x40px, needs 48px on both sides, or is big enough, but detection score 0.40 < 0.55.
• A match below the threshold read like a recognition. match Kirill (0.251) is
  written before the threshold is checked, so a weak candidate looked like a false
  identification — worrying if you drive automations off it. Nothing was ever published
  below match_threshold, but the log did not say so. It now ends in either published or
  below threshold 0.45, NOT published — review queue.
• The occasional go2rtc miss is retried once. Single failures were seen in two
  independent setups while the requests before and after went through; go2rtc has to grab a
  frame from a running stream, and a request landing between keyframes comes back empty.
  Dropping an event over that is a waste when a second attempt costs a fraction of a second.