Skip to content
AI.info

The Pulse

Xailient claims top NIST accuracy among faster face-matching models

Xailient says its face-recognition model had the highest Border-Kiosk accuracy among 106 developers meeting a NIST speed threshold. The company also reported results for a separate live-subject test and compared its model's resource use with competitors.

Xailient claims top NIST accuracy among faster face-matching models

Xailient says its face-recognition model scored the highest accuracy in a NIST Border-Kiosk test among 106 developers whose face-template generation time was no more than 500 milliseconds. At a false match rate of 1 in 100,000, the model correctly matched 95.4% of genuine image pairs in that test, according to the company's release.

Source prnewswire.com

Xailient announced the results on Nov. 17. The company said NIST tested the same model, identified as xailient_001, that it deploys on edge hardware with approximately 1 TOPS of neural processing unit compute.

What the test measured

The company describes the Border-Kiosk test as the only static-image NIST 1:1 test involving non-cooperative subjects. It compares kiosk images with border-crossing images, such as those taken as someone walks through a gate or stands before a border agent.

According to Xailient, subjects use a touchscreen positioned below the camera, creating difficult viewing angles that can include downward facial pitch exceeding 45 degrees. The comparison therefore uses two unposed images.

At a stricter false match rate of 1 in 1,000,000, the model correctly matched the equivalent of 94.0% of genuine Border-Kiosk pairs, the company said. On the easier Border-Border test, Xailient reported a 0.55% false non-match rate at the same false match rate, equivalent to approximately 99.45% genuine-match accuracy.

Speed and resource use

The model generated face templates in 497 milliseconds in NIST testing, according to Xailient. The release puts the on-device benchmark limits at 1 GB of RAM and a 500-millisecond response time, in contrast with server-based or hub-based face recognition, which it says has no fixed limit.

In a comparison with six major competitors, Xailient said its model was 1.6 times faster and produced 12% fewer missed matches than close competitors. It also said the model used 901 MB of memory and 280 MB of configuration data, the lowest figures in that comparison set.

The nearest competitor, according to the company, required 1.8 times more memory and 1.6 times more configuration data. The largest competitor systems required 4 times more compute, up to 6.7 times more memory and 5.7 times more configuration data than Xailient's model, the release said.

From image pairs to live subjects

In a separate independent ISO live-subject test, 250 diverse, moving subjects were matched in real time against a gallery of 10,000 registered candidates. The company reported 99.5% successful recognition with zero false matches in that test.

NIST's FRTE 1:1 evaluation examines one image pair at a time. Xailient says its production system instead evaluates multiple frames as a subject approaches or interacts with a device, combining repeated observations with image-quality gating, matching thresholds and application-level verification logic.

Mallika Patel, Xailient's VP of Product, said matching on the camera avoids collecting bystander data and recognizes only opt-in, officially registered subjects. The NIST results do not constitute an endorsement by the agency of a particular company, product, service or system.

Source

  1. PR Newswire

Related stories

Explore

More articles