Performance Evaluation of YOLOv8 and RetinaFace for Real-Time Face Detection on Edge Devices

Abstract

The demand for efficient and accurate face detection in real-time applications has increased significantly with the rise of Internet of Things (IoT) and embedded vision systems. Deploying deep learning models on low-power edge devices such as the Raspberry Pi 4 presents challenges related to limited memory, computation capability, and energy consumption. This study evaluates the performance of two modern deep learning architectures—YOLOv8 and RetinaFace—for real-time face detection on the Raspberry Pi 4 platform. Both models were tested using the WIDER FACE and FDDB datasets under identical hardware and software configurations. Performance metrics including detection accuracy, latency, frames per second (FPS), and power usage were analyzed to assess suitability for edge deployment. Experimental results indicate that YOLOv8 achieves higher inference speed with acceptable accuracy, whereas RetinaFace provides superior detection precision at the cost of increased latency. The comparative analysis highlights the trade-off between real-time responsiveness and detection accuracy, providing insights into selecting an optimal model for resource-constrained environments. The findings can guide researchers and developers in implementing lightweight, efficient, and reliable face detection systems for next-generation edge-AI applications.