jbaruch/face-recognition-calibration-djl
Empirical calibration for DJL face_feature (ArcFace/FaceNet 512-d) embeddings: cosine distance bands, piecewise confidence formula, enrollment quality targets. Replaces the dlib-based jbaruch/face-recognition-calibration tile for Kotlin/JVM pipelines.
81
86%
Does it follow best practices?
Impact
100%
2.17xAverage score across 2 eval scenarios
Passed
No known issues
Evaluation results
100%
53%Face Recognition Confidence Module
Piecewise confidence formula and identity threshold
Piecewise lower bound
0%
100%
Piecewise upper bound
0%
100%
Piecewise linear region
0%
100%
No textbook formula
100%
100%
Identity threshold 0.60
0%
100%
Threshold vs confidence separation
62%
100%
Cosine distance formula
100%
100%
Correct rejects upper bound in comments
25%
100%
Calibration explanation
100%
100%
Correct cosine distance semantics
100%
100%
100%
55%Face Recognition Runtime Diagnostics and Stability
Frame-drop persistence, garbage detection, diagnostic logging
No-face hold duration
16%
100%
Persistence mechanism
100%
100%
Garbage threshold 0.85
0%
100%
Garbage drops detection
75%
100%
Diagnostic logs all enrolled
100%
100%
Diagnostic distance format
0%
100%
True identity threshold in diagnostics
0%
100%
Others threshold in diagnostics
0%
100%
Spread check in diagnostics
0%
100%
Distance semantics correct
87%
100%
Hold constant named clearly
100%
100%
- Evaluated
- Agent
- Claude
- Model
- Claude Sonnet 4.6