Energy Bugs in Object Detection Software on Battery-Powered Devices

Object detection software is widely adopted in edge computing systems such as mobile devices, drones, and autonomous robots. These edge devices are often battery-powered and, hence, confront stringent requirements for efficient computation and energy saving. Energy bugs that cause inefficient software execution pose significant risks to object detection systems running on battery-powered edge devices. However, such risks have not been well discussed in the previous research. This paper aims to investigate potential energy bugs in object detection software running on a battery-powered device. First, we searched for reports of known energy bugs from internet forums and bug-tracking systems for YOLOv5 and Mask R-CNN (Mask Region-based Convolutional Neural Network). Although we did not find direct mentions of energy bugs, some reports hinted at energy inefficiencies. Consequently, we conducted experiments to assess the impact of energy bugs by monitoring resource utilization and battery consumption of an object detection system running YOLOv5 on a battery-powered Raspberry Pi. Our experimental results clearly show increased power consumption caused by the omission of the input image size setting and improper deployment of the unquantized model, which can be regarded as potential energy bugs in object detection software.