Edge Deployment and Benchmarking of Beet-Root Detection Models
ORIGINAL / RootDown: Deploying and Benchmarking Models for Beet-Root Detection on Edge Devices
This study presents an end-to-end image pipeline for detecting partially occluded sugar beets in loading scenarios, introduces a dedicated benchmark and an automatic annotation strategy using SAM 3, compares models trained on manual versus automatically generated labels, and benchmarks multiple inference configurations including edge devices under realistic constraints of precision, inference time, and power consumption.
01 ABSTRACT
To address annotation and detection challenges caused by overlapping and occluded sugar beets during post-harvest quality inspection, the authors propose an end-to-end image detection pipeline and construct a dedicated benchmark. They use the SAM 3 foundation model to automatically generate bounding box labels, compare models trained on manual and automatic labels, and evaluate edge deployment configurations including a Raspberry Pi 5 with a Hailo accelerator. Results show that foundation-model labels can substitute for a substantial share of manual annotation at a modest accuracy cost, and that the edge device enables high-performance, low-power inference.
02 KEY FINDINGS
- Proposes an end-to-end image pipeline for detecting partially occluded sugar beets and releases a dedicated benchmark.
- Uses SAM 3 to automatically generate bounding box labels, which can replace a substantial share of manual annotation with a modest accuracy drop.
- A Raspberry Pi 5 with a Hailo accelerator achieves high-performance, low-power inference under realistic constraints.
AI GENERATED SUMMARY / DISCOVERED BY ARXIV CS.LG