Delivering Augmented Reality to the Edge: An Approach Toward Object Recognition through the In-Network Computing Paradigm

Abstract

The new advancements on cellular network technology leveraged the 5G beyond a mobile broadband enhancement, enabling major breakthroughs in a wide scope of applications requiring high data throughput and low latency such as Augmented Reality (AR). AR applications promise to be a transformative technology for several fields in the upcoming years, by providing new business opportunities as well as new research trends. In order to deliver such market transformation, AR-based applications must be able to accurately integrate physical and virtual objects in real-time, and for such, it relies on computing-intensive machine learning algorithms. Since mobile devices tend to be computationally and energy-constrained, their capability to provide a synchronous interaction between the real environment, the user, and the visual augmentations is limited. Thus, in this research proposal we propose a method to offload such compute-intensive tasks to programmable network devices at the network edge. By addressing this issue we would benefit both end-users and service providers, as it would diminish the application's end-to-end latency and reduce operating costs, respectively. The solution developed here should be able to run machine learning algorithms on programmable devices, such as switches and FPGAs, both for training and for inference of visual objects.