On the Resource Consumption of 2D Human Pose Estimation Services in MEC-enabled Networks

In this work, we overview key state-of-the-art models and algorithms for 2D Human Pose Estimation (HPE) and provide a detailed discussion on their key features, such as backbone, training dataset and key innovations. Our primary contribution is a comprehensive experimental study on the resource consumption of state-of-the-art HPE models assuming a fully containerized service environment residing at the networked edge/cloud continuum, which is constructed in the form of a Multi-Access Edge Computing (MEC) service compatible with the ETSI MEC architecture. The study serves as a benchmark framework for estimating the offered network, CPU, memory and GPU workload of relevant MEC applications while it also opens the road ahead for service-tailored MEC host association of users that seek to offload their edge-computing tasks in the networked edge/cloud continuum.