Selective Cloud Offloading for Accurate and Efficient Object Detection

Abstract

High-accuracy object detection on resource-constrained devices is essential for many applications including autonomous systems, agriculture, and mobile computing. However, deploying high-performance object detection models on these devices is impractical due to computational limitations, and transmitting and processing all data on a much more powerful remote server running significantly more complex and accurate models, known as full cloud offloading, incurs high latency and cost. This thesis proposes a selective cloud offloading framework that balances prediction accuracy and processing cost. A lightweight edge model makes initial predictions using conformal prediction to quantify uncertainty. Only high-uncertainty regions are offloaded to the cloud for refinement by more powerful models. To further optimize efficiency, multiple uncertain regions are merged into a single image before offloading, reducing transmission and processing costs. The system is evaluated on real datasets, demonstrating substantial accuracy improvements with minimal additional overhead.