Optimizing Edge Offloading Decisions for Object Detection

TL;DR

A reward metric designed to quantify potential accuracy improvements from offloading individual images is introduced, and an efficient approach to make offloading decisions by estimating this reward based only on local detection results is proposed.

Abstract

Recent advances in machine learning and hardware have produced embedded devices capable of performing real-time object detection with commendable accuracy. We consider a scenario in which embedded devices rely on an onboard object detector, but have the option to offload detection to a more powerful edge server when local accuracy is deemed too low. Resource constraints, however, limit the number of images that can be offloaded to the edge. Our goal is to identify which images to offload to maximize overall detection accuracy under those constraints. To that end, the paper introduces a reward metric designed to quantify potential accuracy improvements from offloading individual images, and proposes an efficient approach to make offloading decisions by estimating this reward based only on local detection results. The approach is computationally frugal enough to run on embedded devices, and empirical findings indicate that it outperforms existing alternatives in improving detection accuracy even when the fraction of offloaded images is small.

Figures (13)

• Figure 1: Object detection results that suffer from: (a) Classification error, where a cat is mis-classified as a bird. (b) Localization error, where the predicted location of the surfboard deviates from the ground truth. (c) Background error, where a hallucinated bicycle is detected. (d) Missed ground truth error, where the knife is not detected.
• Figure 2: Images on which the weak detector’s (YOLOv5n) prediction (Top) is considered to be (a) better than, (b) the same with, or (c) worse than the strong detector’s (YOLOv5m) prediction (Bottom).
• Figure 3: Object classes frequency in (a) representative image, and (b) across the Pascal VOC dataset everingham2010pascal.
• Figure 4: Image offloading pipeline. Each image $i$ is first processed by the weak detector $w$ on the local device, based on whose detection results $h_{i,w}$ the reward estimator makes a prediction $e(h_{i, w})$ on the offloading reward. The system decides whether to offload the image to the strong detector $s$ on the edge by comparing the reward estimate with a threshold $T_i$.
• Figure 5: Overall mAP with ORIC and ORI as the offloading reward metrics. ORI assumes $|\mathcal{E}|=0$, while ORIC's performance varies as a function of the context size $|\mathcal{E}|$.