Emdedded Large Scale Face Recognition in the Wild

Cameras have emerged as one of the most important tools to bring ubiquity to the Internet of Things (IoT) since its beginning and can be used to improve contextual accuracy substantially by using effective face recognition technology. The recent literature suggests that there is a large accuracy gap between today’s publicly available techniques and state-of-the-art private face recognition systems. This paper aims at bridging this gap by presenting the results of utilizing the OpenFace public library to perform face recognition in the wild. The key focus of this paper is to present a mechanism that guarantees higher accuracy with low training and less prediction time. In order to avoid the problem of reduced accuracy with the increase in the number of samples, this paper presents a strategy to keep up with low training time and high accuracy in comparison with a range of different recognition/classification techniques. A collection of large data samples of 245 different human faces was utilized to conduct this study, where the training was performed gradually starting from 5 classes to 245 distinct classes, incorporating a total of 7273 images to achieve reasonable accuracy. This paper also includes the details on how the classification was performed on live stream using webcams, where training and classification were done using the NVIDIA Jetson nano platform.