Collaborative Intelligence Framework with Skip-Decoder for Balanced Edge Power and Cloud Cost in Object Detection

This study proposes a cost-effective distributed inference framework for real-time object detection in edge environments. Due to the resource constraints of edge devices, Col-laborative Intelligence (CI) has become increasingly important in enabling efficient distributed inference between cloud and edge. This study leverages the YOLOv7 model to analyze inference latency between edge devices and the server, and presents strategies to address issues related to the power consumption of edge devices and the costs associated with cloud usage. Additionally, a feature map recovery technique called Skip-decoder is introduced to reduce network latency. This novel CI model architecture aims to optimize the use of edge resources, thereby reducing cloud expenses and enhancing real-time object detection performance.