Skip to content

Cross-Quality Face Matching

Definition

Cross-quality face matching addresses the challenge of comparing a high-resolution selfie with a low-quality ID photo. ID document photos are often small (100-200px), printed, faded, and captured years ago — creating a significant quality gap with the live selfie.


The Quality Gap

Attribute ID Document Photo Live Selfie
Resolution 100-200px face 300-800px face
Source Printed on card, scanned Direct camera capture
Artifacts Print dots, moiré, hologram reflections Camera noise, compression
Color Faded, color-shifted True color
Lighting Studio flash (original), ambient (scan) Variable ambient
Compression Multiple generations of compression Single JPEG

AdaFace — Quality-Adaptive Recognition

Aspect Details
Key insight Image quality affects embedding reliability — hard samples from low quality should have less gradient impact
Quality proxy Feature norm (‖f‖) correlates with image quality
Adaptive margin High-quality images get harder margin (more discriminative), low-quality get softer margin (more forgiving)
Result Significantly better cross-quality matching than ArcFace

AdaFace vs ArcFace on Quality-Degraded Benchmarks

Benchmark ArcFace-R100 AdaFace-R100 Improvement
IJB-C (TAR@FAR=1e-4) 96.02% 97.39% +1.37%
IJB-S (Surveillance) 58.29% 63.41% +5.12%
TinyFace 63.89% 72.02% +8.13%

The improvement is most dramatic on low-quality benchmarks (TinyFace, IJB-S).


Other Quality-Aware Approaches

Method Approach
SER-FIQ Estimate quality from embedding robustness (self-supervised)
MagFace Embedding magnitude indicates quality — penalize low-quality equally
FaceQnet Dedicated quality estimation network trained on verification performance
Super-resolution Enhance ID photo resolution before embedding (GFPGAN, CodeFormer)

Key Takeaways

Summary

  • Cross-quality matching is a core eKYC challenge — ID photos are 3-10x lower quality than selfies
  • AdaFace specifically solves this with quality-adaptive margin — best choice for eKYC
  • Super-resolution (GFPGAN) can enhance ID photos but adds computation and may introduce artifacts
  • Quality-aware threshold adjustment provides additional robustness