Revisiting Face Recognition for Monozygotic Twins: The Celeb Twins Test Set
Michael Zang, Haiyu Wu, Mrinal Sharma, Kevin W. Bowyer
cs.CV · cs.AI
Quick verdict
New 21k-pair celebrity-twins benchmark CTTS-80; experiments show CNN matchers ignore skin marks. Valuable but single-matcher, no significance tests.
Summary
The paper introduces CTTS-80 (Celeb Twins Test Set), a web-scraped face verification benchmark of 80 sets of monozygotic celebrity twins, with 132 positive and 132 negative pairs per twin set organized into twin-set-disjoint 10 folds (21,120 pairs total). The authors state it is the only twins test set carrying metadata for distinguishing skin marks (43 sets) and probable mirror twins (7 sets, opposite handedness). Benchmarking four CNN matchers (ArcFace, AdaFace, UniFace, MagFace), each trained on three standard sets (twelve models total), all reach >=97% on LFW-style benchmarks but only 73-77% on CTTS-80, the best being ArcFace/MS1MV2 at 76.50+/-3.42%. Using that model, the authors erase the distinguishing skin mark in images of three twin sets (e.g., Tamera Mowry) and find distance distributions essentially unchanged, with original-vs-erased same-person distances clustering near zero - direct evidence that matchers ignore skin marks. An asymmetry-aware ArcFace retrained without horizontal-flip augmentation scores 78.58%+/-3.7, no improvement. Pilot experiments with Grok, ChatGPT, and Gemini show generated imagined twins are 'literally identical' (wholly overlapped distributions) or refused, so generative-AI twins training data remains unsolved. Limitations: detailed analyses use a single matcher, no formal statistical tests, erasure on three example sets only, and possible train/test identity overlap is acknowledged but unquantified.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 2
● low 3
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- CTTS-80 is the largest and only metadata-rich (skin marks, mirror handedness) twins verification test set, organized like LFW-style 10-fold protocols with twin-set-disjoint folds.
- Current deep CNN matchers achieve only about 73-77% accuracy on CTTS-80 (best: ArcFace/MS1MV2, 76.50+/-3.42%) versus >=97% on the traditional five benchmarks.
- Current matchers do not encode visible skin marks: erasing them leaves distance distributions essentially unchanged, and original-vs-erased same-person distances cluster near zero.
- Retraining without horizontal-flip augmentation (asymmetry-aware) does not significantly improve twins accuracy (78.58%+/-3.7), and mirror-twin distribution overlap changes are mixed (Table 2).
- Current GenAI tools cannot produce plausible imagined-MZ-twin training data: Grok and ChatGPT yield literally identical twins (wholly overlapped distributions) and Gemini declines to generate.
Strengths
Largest, metadata-rich twins verification benchmark
CTTS-80 provides 21,120 pairs across 80 twin sets (~3.5x ND-Twins' 6,000) with unique metadata: 43 sets with distinguishing skin marks and 7 sets with opposite handedness (mirror twins); the twin-set-disjoint 10-fold design (Sec. 5) prevents a twin set's pairs from influencing its own threshold.
Clean causal probe of skin-mark usage
Erasing the distinguishing mark leaves distance distributions essentially unchanged for three twin sets (Mowry, Hassan, Kaczynski; Figs. 10, 12, 14), and original-vs-erased same-person distances cluster near zero (e.g., Fig. 10c) — an interpretable intervention showing marks carry negligible embedding weight.
Honest variance reporting
Timely negative result on GenAI twins
Useful literature synthesis
Weaknesses
Single-matcher generalization
No formal statistical inference
Erasure pipeline unspecified
Unquantified train/test overlap
Anecdotal GenAI probe
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper delivers three linked contributions: (1) CTTS-80, a web-scraped face verification benchmark of 80 sets of monozygotic (MZ) celebrity twins — 264 pairs per set (132 positive, 132 negative sampled from 144), organized in twin-set-disjoint 10-fold cross-validation (8 sets per fold, 2,112 pairs/fold, 21,120 total, ~3.5x ND-Twins' 6,000 pairs); (2) an empirical test of whether current deep CNN matchers exploit skin marks or mirror asymmetry, the two cues emphasized in the pre-deep-learning twins literature; (3) a feasibility probe of generative AI for creating imagined-twin training data. Central claims: the best of 12 public matcher instances reaches 76.50+/-3.42% on CTTS-80 versus >=97% on the 'Trad 5' benchmarks; visible skin marks have essentially no effect on embeddings (erasure experiments, Figs. 10/12/14); an asymmetry-aware retrain (no horizontal-flip augmentation) yields 78.58%+/-3.7, no significant gain; and Grok/ChatGPT imagined twins are 'literally identical' while Gemini declines to generate.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
How accurate is face recognition at telling identical twins apart?
On the new CTTS-80 benchmark of 80 celebrity twin sets, CNN matchers that exceed 97% on standard benchmarks like LFW manage only about 73-77% verification accuracy. A 2022 NIST study similarly found that at a threshold set for a 1-in-10,000 false match rate on non-twins, most algorithms falsely accepted a twin about 98-99% of the time, with the best at 0.475.
What is the Celeb Twins Test Set (CTTS-80)?
A web-scraped face verification benchmark with image pairs for 80 sets of monozygotic celebrity twins: 21,120 pairs organized in ten folds, with 132 same-person and 132 different-person pairs per twin set. It uniquely includes metadata on distinguishing skin marks (43 sets) and probable mirror twins identified by opposite handedness (7 sets), and is available for non-commercial research via the authors' GitHub (mzang20/CTTS).
Do face recognition systems use moles or freckles to identify people?
According to this paper's experiments, no: erasing a distinguishing skin mark from images left the matcher's distance distributions essentially unchanged, and same-person distances between original and mark-erased images clustered near zero, showing the mark has almost no effect on the computed embedding. The paper attributes this to twins being rare in training data and to horizontal-flip augmentation.
Can AI image generators create fake identical twins for training face recognition?
Not yet, per this paper. Grok and ChatGPT generated imagined twin pairs whose faces were literally identical — their same-person and different-person distance distributions wholly overlapped, unlike real twins — while Gemini declined the request. The authors argue generators lack knowledge of how real monozygotic twins' faces naturally differ.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure images (Figs. 1-17) cannot be assessed from text extraction; judgments about image content (skin-mark visibility at 112x112, plausibility of AI-generated faces, distribution shapes) rest on captions and prose claims only.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.