Comparison of Split Computing Scenarios for Object Detection

Abstract

This paper presents a detailed comparison of three split computing scenarios. As a toy task for this comparison, we choose object detection - a very common task for embedded systems and real-world scenarios. As baseline model, we use the YOLOv8 model, which is already optimized for real-time computation. We show that by the addition of a simple bottleneck into this networks, we can decrease the computation time by 80% comparing with the version without the bottleneck with only a negligible decrease in the detector's performance. Our code is available at https://github.com/YvanG/split_computing/tree/master.