Skip to content

ofiqpy

A faithful Python port of OFIQ v1.1.0 — the ISO/IEC 29794-5 face image quality reference implementation from the German Federal Office for Information Security (BSI).

Unlike an ISO-inspired reimplementation, ofiqpy reuses OFIQ's own model files and reproduces its exact .cpp algorithms, targeting per-component agreement with the OFIQ reference to within ±1 quality point — the ISO/IEC 29794-5 Annex A conformance criterion.

Conformance at a glance

Validated on 1,000+ real CelebA images against a live OFIQSampleApp v1.1.0 run:

  • 27 of 28 components fully conformant (±1 on every image); most are bit-exact (Δ=0).
  • The handful of per-image residuals are numerical boundaries of discrete/learned models (the RTrees vote count, an AdaBoost score, a sigmoid round) — not algorithm gaps.
  • ~99.99% of all component-image pairs within ISO ±1.

See Conformance for the full methodology and per-component figures.

What it computes

All 27 ISO/IEC 29794-5 quality components plus the unified quality score:

  • Capture / pixel: background & illumination uniformity, luminance mean/variance, under/over-exposure, dynamic range, sharpness, compression artifacts, natural colour.
  • Subject / geometry: single-face, eyes open/visible, mouth closed/occluded, face occlusion, no head coverings, expression neutrality, inter-eye distance, head size, crop margins (leftward/rightward/above/below), head pose (yaw/pitch/roll).
  • Unified: MagFace-based unified quality score.

Design

  • Same weights — models are loaded directly from an OFIQ install (see Installation); nothing is re-trained.
  • Same math — detection, ADNet landmarks, 5-point similarity alignment, the landmarked-region mask, and every measure are ported line-faithfully from OFIQ's C++.
  • Same versions — pinned to OpenCV 4.5.5 + onnxruntime 1.18.1 (OFIQ's toolchain) for bit-faithful model inference.
  • Gated, not asserted — a harness runs live OFIQ and checks |port − OFIQ| ≤ 1 per image.
from ofiqpy import assess
scores = assess("face.jpg")          # {component: (raw, scalar)}
print(scores["UnifiedQualityScore"]) # (raw_magnitude, 0-100 scalar)