Better Drinking Through GPU-Powered Facial Recognition at GTC

by Tony Kontzer

We all know the cameras are out there. In banks and government buildings. In trains stations and outside warehouses. They’re capturing images of people’s faces to help  prevent nefarious activity.

At the GPU Technology Conference, they’re capturing your drink order.

Herta Security, which makes facial recognition technology that powers security efforts at facilities worldwide, has teamed up with NVIDIA for a unique GTC show floor demonstration at the aptly named YouBar.

Here’s how it works:

Walk up to the registration counter next to the YouBar,  get your face recognized and place your drink order. Then, at your leisure, you can approach the YouBar. Your face will get identified on monitors behind the bar. Tipped off, the barkeep greets you by name and serves up your favorite drink.

Cursory testing by your correspondent suggests the system works. We probably don’t have to point out that this was a popular GTC happy hour destination.

But before anyone starts making plans to open a facial recognition bar, they should know that Herta has no plans to bring this particular application of their technology to market. Rather, Herta CEO Javier Rodriguez Saeta sees the YouBar as a fun way to familiarize GTC attendees with the technology.

But Saeta, who was at the YouBar Tuesday evening answering questions, did say the company is eyeing potential markets outside its strongholds in security and law enforcement.

More Than Just a Face

Earlier in the day, YouBar’s director of research, Carles Fernandez Tena, said as much during a well-attended GTC session. He said the technology’s ability to recognize a face, identify its gender and ethnicity, theorize about its age and analyze its expression could have significant implications for marketers looking to provide more personalized service.

The big opportunity for Herta could be retail. That’s why Herta is tapping GPU technology more than to improve its speed and accuracy, focusing those refinements on video decoding.

Tena said that decoding on CPUs represents a “major obstacle” to achieving the desired speed and accuracy, which is why the company has been experimenting with decoding on NVIDIA’s new Jetson TX1 embedded computing platform.

So far, doing so has slashed decoding times by two-thirds. It has also enabled Herta to more effectively use a process known as “frontalization,” in which images of faces that are angled or tilted can be reoriented to make them appear to be looking straight ahead.

Tena said this is boosting accuracy from 96 percent to 98 percent, which he described as a “huge increase” that’s much more substantial than the 2 percentage point improvement implies.

Naturally, this has boosted Herta’s enthusiasm for the Jetson TX1.

“We want the TX1 to do everything,” Tena said.

Everything, that is, except make your cocktail.