YuNet — Ultra-High-Performance Face Detection in OpenCV — a good solution for real-time POC, Demo, face applications.

Face Detection is an absolute required module for face recognition, face classification, face attribute estimation, etc. Recently, when making a quick, effortless, and high-performance demo for my SaaS service, I, once again, had a chance to experience several face detection in the world.

Let’s take a look into the family of Face Detection.

Face detection models [https://paperswithcode.com/]

Models on the green path achieved the top AP at their periods. However, employing them in a practical application requires effort. Some models are not published because of the authors’ privacy. Some models are implemented in the inaccessible framework for data scientists or non-ML engineers.

Moreover, along with accuracy, speed is the most crucial aspect of a Proof of concept (POC) or a demo. Customers, indeed, do not want to see any lags before adding our service to their solution. Plus, deploying on CPUs is the proper approach to save running costs instead of inferring by GPUs. RetinaFace and TinaFace (even MTCNN and Dlib) are not sufficiently swift to fulfill the mentioned requirement. I have verified the operating speed of such models. It is about 0.3–0.4 seconds for one model to predict the results on a frame of 640x480 resolution. It means you only process…