A Metric for Evaluating Reference-Free Demorphing Methods

Research Goal

As reference-free demorphers proliferated, there was no agreed way to evaluate them — and the popular metrics are actively misleading. This work diagnoses why and proposes a metric that faithfully measures how well a demorpher recovers the true constituent identities.

Why existing metrics fail

• TMR & Restoration Accuracy reward cheating. A morph is, by construction, biometrically close to both its constituents — so a trivial demorpher that just outputs the morph twice scores ~100% TMR/RA. The metric can’t tell real demorphing from regurgitation.
• SSIM/PSNR ignore identity. Operating purely in RGB pixel space, they can rate a face as structurally closer to a different identity than to a noisy copy of itself — nonsensical for a biometric task.

The proposed metric

Biometrically cross-weighted IQA — BW(iqa): weight each output↔ground-truth image-quality score (SSIM or PSNR) by the biometric match score between the same pair, and take the max over the two possible (unordered) pairings. This couples structural fidelity in pixel space with identity fidelity in feature space, so scores track what a human sees.

Key results

• Benchmarks three demorphers (Facial Demorphing, SDeMorph, IPD) across six datasets and two matchers (AdaFace, ArcFace).
• Under BW-IQA, IPD ranks best (BW-SSIM 0.23 / BW-PSNR 6.0), ahead of Facial Demorphing (0.17 / 3.84) and SDeMorph (0.14 / 3.62) — consistent with visual inspection, unlike TMR/RA which crown the trivial solution.

Resources

Results

Citation

If you use this work, please cite:

@inproceedings{shukla2025metric,
  title={Metric for Evaluating Performance of Reference-Free Demorphing Methods},
  author={Shukla, Nitish and Ross, Arun},
  booktitle={IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACVW)},
  pages={1670--1676},
  year={2025}
}