Obviously, the revolution in deep neural network-based algorithms for computer vision tasks, including face recognition, is changing the biometrics market today.
Obviously, the revolution in deep neural network-based algorithms for computer vision tasks, including face recognition, is changing the biometrics market today. We have come a long way from narrow application scope with unprecedented investments needed to implement high-quality face recognition in national security areas to the mass commercial adoption of this technology.
However, according to Alexander Khanin, Founder & CEO of
VisionLabs, there are several factors to consider before deciding which solution provider to go with. Many facial recognition solution providers are yet to migrate to advanced artificial intelligence-based solutions and customers have to be aware of what to know before selecting the product.
“We have to acknowledge the fact that even though there are hundreds of face recognition market players today, there are distinct features visible to any professional that both can confirm the maturity of the technology offered by the face recognition engine supplier as well as make the expectations from the working system in real life conditions match marketing materials,” Khanin said, adding five key questions to consider.
1. How many facial points are used for face recognition?
If the answer is 10, 50, 80, 90 etc., the solution uses an algorithm from the previous century. State-of-the-art face recognition technology does not use facial points for recognition nor do they measure the distance between the eyes: the easiest trick to find this out during a demo is to walk into the camera surveillance zone with your eyes closed and the older systems will see nothing. The same goes for glasses – they cover so many key facial points, making older systems fail, while deep neural networks-based algorithms perform well.
“Remember that justifications like, ‘there is no need to use deep neural networks for the full face recognition pipeline’, is simply an excuse to not being top market computer vision and machine learning experts, and, that also leads to the next tip,” Khanin said. “Modern face recognition platforms have both face detection and face recognition algorithms 100 percent based on deep neural networks.”
2. What machine learning inference framework do you use?
The answer in most cases will be the open source framework, since it is very hard to make your own. Therefore, companies go for deep neural network-based algorithms just partially in their products – they require tremendous computing power and are slow in the wrong hands.
3. For uncontrolled face recognition, what camera do you need?
If the answer is special cameras costing US$5,000 to $7,000, this is also the legacy of the previous century, according to Khanin. Modern face recognition algorithms based on deep neural networks are fine with almost any IP camera in $400-1000 price range.
4. Do you provide on-premise solution?
This is a must have for the airport terminal. One tricky question here is that face recognition providers will start selling you a system that will self-machine-learn on premise.
“Ask yourself one question: do you really know what is the process of training of CNN [cellular neural network]-based algorithms?” Khanin said. “Do you know what kind of resources and compute power it requires? So, marketing is not science.”
5. Have you run the tests?
Run systems providers’ products in parallel for the PoC. Remember, you do not have to pay 5, 10 or 20 providers for this. Good-quality products will not require months of implementation for the PoC – it just takes hours. Good companies will not make you wait for weeks to receive the results of testing, doing something at their office and then sending you some kind of a chart. Everything is done automatically now.