RealNetworks’ investment in machine learning tech has yielded many diverse results. Last year, it moved the company into the facial recognition space. In an exclusive video, RealNetworks’ Mike Vance talks about the SAFR platform…

Most people would know RealNetworks as a streaming music and video specialist. Many MEF insiders might think of RealNetworks as a company that can classify text messages at scale.

But would anyone associate RealNetworks with facial recognition?

North American school pupils might. Why? Because in July 2018, RealNetworks launched a facial recognition platform called SAFR. On day one, it provided the technology free to every elementary, middle, and high school in America and Canada.

It might seem like a jump for RealNetworks to move into the facial recognition space. Actually the move is quite logical given the company’s history of compressing video and  – more recently – investing in machine learning.

Mike Vance, senior director of product management at RealNetworks, explains: “If you look back at the 25 years of the company, we have a history of compressing and streaming video through a very small pipe – and doing analytics on the content and context of the files. So SAFR really came out of that expertise.”

According to Vance, machine learning has led to huge advances in facial recognition. “We start with the equivalent of an untrained human brain,'” he says. “When a child is born, a baby knows ‘mom or not mom, food source or not food source’. A modern AI system is trained the same way. You feed it data, and tell it ‘you got that right’ or ‘you got that wrong’. Over time, it learns and it creates its own rules.

  If you look back at the 25 years of the company, we have a history of compressing and streaming video through a very small pipe – and doing analytics on the content and context of the files. So SAFR really came out of that expertise”

“The amount of data that’s out there means that people can train models on larger data sets… this has made it possible to achieve accuracy levels that were unachievable previously.”

Of course, there is far more to facial recognition than the underlying technology.

The topic brings with it profoundly important questions around privacy. People are naturally concerned about improper use and the sharing of facial profiles.

Vance says the SAFR system addresses these concerns. “All the data is owned by the customer,” he says. “We don’t syndicate it or have one ID across multiple customers.”

He adds that the platform can be configured to confirm a person’s identity locally, without any need to send a file to the cloud. “In some case people want everything to happen in their own environment. We can operate with zero internet connection.”

SAFR also avoids some of the issues that have arisen around ethnic profiling. There have been questions raised about the way some systems categorise people by race, for example.

Vance says SAFR nullifies these questions by avoiding ethnicity as a data set completely. “One of the big challenges in facial recognition is optics. Lighter faces in lighter backgrounds get washed out, while darker faces in darker environment can be harder to detect. So one thing you can do is start with a geographical and ethically diverse data set.

“However we don’t use skin tone or gender to train the model. We only use geometry. So the system is just looking at vectors and angles. And we are at 99.86 per cent accurate for ‘camera unaware’ faces in the wild.”

RealNetworks has successfully exported its SAFR tech to partners in Japan and Latin America. It is targeting use cases across retail, health, telecom, education, government, and workspace markets.