Prostock-studio - stock.adobe.co
Intel introduced a facial recognition component to its Intel RealSense product line that is designed to provide fast and accurate facial recognition and authentication.
Launched Jan. 6, RealSense ID includes an active depth-sensing camera with a dedicated system-on-chip containing Intel's specialized facial authentication neural network.
The product comes as advances in computer vision and neural networks have made facial recognition and authentication more sophisticated, said Alan Pelz-Sharpe, founder and principal analyst at Deep Analysis.
"There have been advances over the past several years and the results can be highly accurate," he noted.
Picture or person
With the active depth sensor, RealSense ID can tell the difference between a real, present human face and a video or static image of that same face, according to Intel.
The algorithm can adapt to facial changes over time, identifying users even if they change their hairstyles or put on the sunglasses, said Joel Hagberg, head of product management and marketing at Intel RealSense, during a Jan. 5 press briefing.
The product can also accurately identify users if they wear a mask, an important feature during the COVID-19 pandemic; however, the user must have their nose exposed for the algorithm to work.
Alan Pelz-SharpeFounder and principal analyst, Deep Analysis
Although facial recognition technologies are growing in popularity, masks have interfered with their adoption and accuracy, Pelz-Sharpe noted.
"To be fair, not something anyone planned for, but not the best circumstances for facial recognition software vendors," he said.
RealSense ID stores all its data securely within the device, and users can get started with it after a short setup time.
"We don't store your images, it's basically intrinsic data," Hagberg said. "We're not storing video images and things that can expose a user's information."
The data "is mathematically important to our algorithm, but not useful to anyone outside of our algorithm," he noted.
For users that want to integrate the product into a device that already contains a camera, such as an ATM or a self-service kiosk, Intel is also selling a small module containing just the neural network.
"We've built a lot of intelligence and the algorithm into this very small module with an onboard processor. We've got a lot of power and performance packed into a very small device," Hagberg said.
Racial bias an issue
Many facial recognition technologies have struggled to accurately identify people of color, and
several vendors, including Amazon, Microsoft and IBM, stopped the sales of facial recognition software to law enforcement agencies last year for just that reason.
The reason, in part, is that companies and researchers haven't trained their models on enough data that represents people of color. That's getting better, Pelz-Sharpe said, but it still has a long way to go.
"From a purely technical side of things, one can argue that the gap is closing, and it will improve over time," he said. "From a cultural, ethical and possible legal side of things, you need to tread very cautiously.
Hagberg said that Intel has done "extensive data collection of all ethnicities, from Asian, Europe, to Middle-East Africa."
"We were very careful to have all ethnicities covered," he said.