All state-of-the-art face recognition algorithms today are based on deep neural networks and our algorithms are no exception to that. While CNN ensembles trained on large-scale databases of face captures deliver superior quality throughout many stages of face recognition pipeline, they lead to much higher compute power requirements to deliver real time stable performance. How does VisionLabs deliver great results? - We choose NVIDIA.
Face recognition algorithms training & inference
Now it takes us just several days for precise training our algorithms on hundreds of millions facial images when we use NVIDIA GPUs compared to previously months with common CPUs. Using NVIDIA GPUs for inference brings visual recognition systems performance to the new level.
Face detection algorithm
Even single NVIDIA Tesla makes multiple faces detection in live video streams 10x faster for both controlled and uncontrolled face capture scenarios, while NVIDIA Jetson delivers superior performance as the part of embedded face recognition systems.
Face descriptor extraction algorithm
Our face descriptors are computed with an assembly of Convolution Neural Networks deﬁned by tens of millions of weights and parameters. Up to 15x speed up boost with NVIDIA GPUs allows us to deliver solutions with uncompromising high accuracy.
Security video surveillance and access control
Real time city-scale 1:many person identiﬁcation in uncontrolled video surveillance on thousands of camera channels, millions of enrolled people in various watch lists and 1:1 person authentication for access control application scenarios in enterprises of any size.
Banking and finance
Identity fraud prevention by 1:many matching of new incoming face captures with a nationwide database of banks’ Customers PII such as identity document photograph, previously enrolled face photographs as per banks’ internal procedures or virtually any external database.
Accurate gender (99.5%) and age (± 1,5 years) detection with integrated Customer authentication at POS, mobile and WEB platforms of retailers with practically unlimited number of camera channels processing capacity.