Since 1999, Canadian photographer François Brunelle has been chasing a fascinating mystery: Can two complete strangers look like identical twins?
His ongoing photo project “I’m Not a Look-A-like!” features striking black-and-white portraits of unrelated people who resemble each other so much, they could easily pass as twins. Over the years, Brunelle has photographed more than 250 such pairs in 32 cities around the world.
But how does facial recognition technology see these “twins”? We decided to find out.
We ran several of Brunelle’s look-alike portraits through our online facial recognition demo.
The result? The algorithm didn’t find a single match—even in cases where most people would swear it was the same person.
Let’s take a closer look at how the experiment worked — starting with how a face recognition algorithm decides: “same person” or “no match”.
How Face Recognition Knows Who’s Who
To make a decision on whether two faces match, the algorithm relies on a trade-off graph between two types of errors:
False Rejection (FRR):
A real user is wrongly denied.
→ Leads to frustration, drop-offs, failed onboarding.False Acceptance (FAR):
An imposter is wrongly accepted.
→ Leads to security breaches and compliance risks.
These are measured via a similarity score threshold. Tighten the threshold = fewer imposters get through, but more real users get blocked. Loosen it = smoother UX, but higher risk.
The default recommended similarity score threshold for our algorithms is 0.85. At this threshold, the algorithm achieves:
- False Acceptance Rate (FAR): 0.0000009919
- False Rejection Rate (FRR): 0.0075107813
This balance ensures extremely low chances of unauthorized access while maintaining high acceptance for legitimate users.
Now let’s move on to specific examples taken from the website http://www.francoisbrunelle.com/webn/e-project.html
What the Lookalike Analysis Revealed
Example 1.
Score is 17. FAR=0,019458; FRR=0,003017. Verdict: Different people.
Example 2
Score is 23. FAR=0,007506; FRR=0,003363. Verdict: Different people.
Example 3
Score is 14. FAR=0,053261; FRR=0,002691. Verdict: Different people.
And for a clearer visualization — here's a comparison of individuals taken from different lookalike pairs.
Score is 15. FAR=0,034590; FRR=0,002873. Verdict: Different people.
These results show that even when two faces appear strikingly similar, face recognition algorithms can still distinguish between them with high accuracy — far outperforming human judgment in such cases.
But here’s the catch: accuracy isn’t guaranteed. It still hinges on a few critical factors — like quality and diversity of training data, the underlying model architecture, and the real-world conditions.
What Really Impacts Face Recognition Accuracy
Despite impressive results, face recognition performance still depends on several key factors:
Training data quality: The more diverse and extensive the data used to train an algorithm, the better it performs in recognizing different types of faces (varying in age, race, and gender). High-quality and well-balanced datasets significantly boost accuracy.
Model architecture: Modern face recognition algorithms—especially those based on convolutional neural networks (CNNs)—achieve high accuracy thanks to deep learning and the ability to detect subtle facial features. Nearly all leading market players use complex neural architectures for precise facial identification.
Image quality: Just like with human recognition, image clarity is critical. Sharp images with good lighting significantly improve recognition accuracy, while blurred, dark, or partially obscured faces can challenge the system (Tip: Tools like 3DiVi’s QAA can pre-filter low-quality images before they hit your system).
Appearance changes: Contemporary algorithms are capable of recognizing faces despite minor appearance changes (e.g., hairstyle, makeup, or glasses). However, drastic alterations can make recognition more difficult.
Cross-race effects: Just like humans, algorithms may be biased depending on the data they were trained on. If the training set lacks ethnic diversity, algorithms may struggle to recognize underrepresented groups. Still, even accounting for this effect, modern systems make significantly fewer mistakes than humans in similar conditions.
Where to Find Performance Metrics
FRVT (Face Recognition Vendor Test): Tests on various datasets show that in 1:1 verification scenarios (e.g., matching a passport photo to a person), top systems have an error rate of less than 0.01%. In 1:N identification tasks (e.g., searching a face in a database), accuracy remains high but depends on database size and image quality.
Real-time face recognition: In challenging environments (e.g., outdoor surveillance), accuracy can drop. However, advanced algorithms still achieve over 95% accuracy, particularly when improved image processing and adaptive techniques are applied.
Examples of Real-World Systems
Apple Face ID delivers 98–99% accuracy in optimal lighting and angles — impressive for a consumer device.
3DiVi algorithms, benchmarked by NIST, show world-class performance with a False Match Rate (FMR) of 0.000001 and False Non-Match Rate (FNMR) of 0.003 at the default threshold.
Final Thoughts
Modern face recognition algorithms are pushing the boundaries—reaching near-100% accuracy and outperforming humans in controlled settings. But it’s not without limits — bias in training data, false match rates, and poor image quality can still get in the way.
Top comments (0)