Sface: the fastest (also powerful) deep learning face recognition model in the world

I published about YuNet — Ultra-High-Performance Face Detection in OpenCV — a good solution for real-time POC, Demo, and face applications. This is for face detection. I’m curious if any deep-learning facial recognition models can run fast and light while achieving the capacity of a huge data-trained model. SFace: Sigmoid-Constrained Hypersphere Loss for Robust Face Recognition answers my question. When you combine SFace and Yunet, you can create a real-time face recognition system that works with any operating system.

I will not explain the process behind SFace’s power, but I strongly advise people to read the technique in the paper. In a nutshell, the benchmark in the research produced astounding results, indicating that SFace outperformed the robust and popular models ArcFace, SphereFace, and CosFace in the public validation dataset: LFW, CFP-FP, AgeDB-30, CPFLW, and CALFW. The speed of SFace is what most impresses me. SFace is now tied to OpenCV, which is extremely useful for designing any operating system in any programming language.

I ran the system through its paces on my MacBook Pro 2019, 1,4 GHz Quad-Core Intel Core i5. Face detection average speed with Yunet is 0.04 seconds/image for continuous operation with an input 720P camera, and SFace consumes roughly 0.01 seconds/face (again, outstanding). This means that SFace can process up to 100 frames per second.