Efficient Face Detector Using Spatial Attention Module in Real-Time Application on an Edge Device

Abstract

The practical application requires a vision-based face detector to work in real-time. The robot application uses the face detection method as the initial process for the face analysis system. During its development, it utilizes an edge device to be used to process sensor information. Jetson Nano is a mini portable computer that is easily synchronized with sensors and actuators. However, traditional detectors can work fast on this device but have low performance for occlusion cases, multiple poses, and small faces. On the other hand, CNN-based detectors that implement deep layers are slow to run on low memory GPU devices. In this work, an efficient real-time face detector using a simple spatial attention module was developed to localize faces rapidly. The proposed architecture consists of the backbone module to efficiently extract features, the light connection module to reduce the size of the detection layer, and multi-scale detection to perform prediction of faces on various scales. As a result, the proposed detector achieves competitive performance from state-of-the-art fast detectors on several benchmark datasets. In addition, this efficient detector can run at 55 frames per second in video graphics array resolution on a Jetson Nano.