Performance analysis of Convolutional Neural Networks on edge devices for Computer Vision tasks

Computer vision is currently applied in an increasing number of technological systems and devices. In many cases, security and privacy constraints, or the need for real-time decision-making, require these tasks to be executed at the edge, where images are acquired. When high performance targets must be met, Convolutional Neural Networks (CNNs) remain the gold standard since, if compared to more recent and complex architectures, they provide a simpler structure that allows for easier implementation and compatibility with different hardware platforms. This paper presents a comparative analysis of the performance of several state-of-the-art CNNs on two edge computing architectures, specifically Jetson Nano and OAK-D-CM4. We considered also the Coral Edge TPU, even if it seems discontinued. The objective is to evaluate the achievable performance and identify the limitations inherent in the available software libraries and hardware. Particular attention is given to the trade-off between high accuracy and fast inference. To this end, two use cases targeting classical Computer Vision tasks, i.e. object detection and face recognition, will be discussed.