Hidden Challenge in Deep-Learning Real-Time Object Detection on Edge Devices

As the need arises for edge intelligence, which combines edge-computing with artificial intelligence, efficient and accurate algorithms are increasingly needed. Edge computing allows AI applications to be executed in real-time, closest to hardware devices/systems, thus reducing latency that could be introduced if traversing the cloud or sent to some storage unit. The aim is the needed real-time information processing, among other benefits. Notwithstanding the advantages of Edge-AI, a challenge that is not generally addressed in many such real-time detection and recognition DL models is the effect of an object rotation on real-time object detection accuracy. In particular, object label uncertainty can be introduced in rotated object detection because the detected shape is not foreseen. The focus of this paper is to highlight how real-time object detection can be affected by rotation, a problem of importance in the field. We selected a particular architecture, Efficientdet, and implemented it on a Raspberry Pi 4 to study this rotational object detection handicap. This paper also aims to increase awareness and tender possible solutions in hopes that future researchers and engineers can design DL models on edge computing devices that will significantly diminish the impacts of an object rotation on its real-time object detection accuracy.