face-recognition-calibration-djl

A Tessl plugin encoding empirical calibration for DJL face_feature embeddings (ArcFace-derived, 512-d L2-normalized, cosine distance) — the Kotlin/JVM equivalent of jbaruch/face-recognition-calibration, which targets dlib's Python face_recognition library instead.

What this plugin provides

Why it exists

The textbook similarity formula conf = 1 - d / tolerance (where tolerance = 0.6) is technically correct and demo-wrong for DJL face_feature distances. A strong recognition at d ≈ 0.30 returns conf = 0.50 — yellow in any 3-band semaphore. The user leans into the camera to "improve the signal" and confidence climbs from 0.50 to 0.67. They look uncertain about the demo.

The piecewise formula in this plugin returns conf = 1.0 at d = 0.30 — solid green where it belongs. Calibrated against the actual distance distribution we measured on hardware: strong matches cluster at 0.18–0.40, rejects above 0.65, with a clear gap between.

Use this plugin if

• You're using DJL face_feature (PyTorch engine, 512-d) for face recognition on Kotlin/JVM.
• A strong-looking recognition still reads as "weak" downstream.
• You're driving a confidence display (semaphore, progress bar, gauge) that needs strong matches to register as strong.

Use jbaruch/face-recognition-calibration instead if you're on Python with dlib's face_recognition library — the constants differ (d ≥ 0.60 → 0.0 for dlib L2 vs d ≥ 0.65 → 0.0 for DJL cosine).

Install

tessl install jbaruch/face-recognition-calibration-djl

License

MIT — see LICENSE.