RootDown: Deploying and Benchmarking Models for Beet-Root Detection on Edge Devices

1 Introduction

• Current visual inspection is limited to truck-bed top layers, and time constraints restrict estimates to a small number of individual beets.Soil settling at the bottom of truck beds further reduces inspection effectiveness.Source: §1 · pp. 1–2
• Robust detection and separation of moderately to highly overlapping conveyor beets is the paper’s first step toward automated visual inspection.The proposed pipeline integrates acquisition, annotation, training, evaluation, and deployment, assessing inference time, precision, and power consumption.Source: §1 · p. 2
• RootDown introduces a benchmark for partially overlapping beets in realistic loading operations and a foundation-model-based strategy to reduce manual labeling effort.The work also examines annotation sources, recording conditions, and camera viewpoints, and measures accuracy, inference time, and power across embedded and general-purpose platforms.Source: §1 · p. 2

2 Related Work

• Classical inspection pipelines showed promise but were often evaluated with controlled backgrounds and fully isolated objects.Prior root-vegetable work includes potato yield estimation, while SemanticSugarBeets focuses on non-overlapping beet segmentation.Source: §2 · p. 3
• Modern detectors perform well in clutter but require large, carefully annotated datasets, while reported produce systems often use fixed lighting and no overlap.Occlusion-aware methods include part-based models, attention, and explicit overlap modeling such as UOAIS.Source: §2 · p. 3
• Foundation vision models offer zero-shot segmentation but are computationally demanding for real-time edge deployment, and annotation bootstrapping can degrade under heavy clutter and occlusion.The paper identifies partially overlapping natural objects, especially root crops in operational settings, as underexplored and rarely evaluated on energy- and compute-constrained platforms.Source: §2 · p. 3
• The paper addresses these gaps through a dedicated benchmark, semi-automatic annotation workflow, and analysis of training and deployment under realistic constraints.Source: §2 · p. 4

3 RootDown Benchmark

• RootDown was created because no publicly available conveyor image dataset existed, using operationally acquired images and both manual and fully automatic annotation methods.Source: §3 · p. 4

3.1 Image Acquisition, Sampling and Meta Annotation

• The benchmark combines 597 selected images from 52 loading sessions over three days with complementary high-resolution images of soil-covered ground beets.Three GoPro cameras captured loader viewpoints; recordings were sampled at 1 fps and stationary-conveyor periods excluded.Source: §3.1 · p. 4
• Annotated metadata covers camera scenario, natural lighting, and recording location; distributions span realistic lighting, while Side is rare because vibration shifted the camera from Back.Scenario and Location are fairly evenly distributed, and diffuse lighting is common during autumn and winter harvest conditions.Source: Meta Parameters · p. 5

3.2 Manual and Automatic Annotation

• SAM 3 with the prompt vegetable recovered approximately 90% of manually identified instances and achieved 46.1% mAP50–95 against manual annotations.Generating annotations for all 597 images took 20 minutes on an NVIDIA RTX 3090 GPU.Source: §3.2 · p. 6

4 Methodology

• Object detection is the pipeline’s first stage, while downstream tasks such as counting, region-based classification, and soil-adhesion estimation are left for future development.Accurate beet localization is intended to support those later tasks.Source: §4 · pp. 6–7

4.1 Benchmark Ablation

• Core-Manual uses only manual labels; Core-Auto replaces training and validation labels with automatic labels, while Ext-Auto adds SAM 3 images and Full-Auto uses all available automatically annotated images.All variants share the same image split, and the test set uses identical manually annotated data.Source: §4.1 · p. 7, Table 1 · p. 7

4.2 Hardware Setup

• The hardware evaluation includes laptop and workstation CPUs, a workstation GPU, and a Raspberry Pi 5 with or without an NPU; Hailo-8L and Hailo-8 provide up to 13 and 26 TOPS, respectively.The hardware overview reports INT8-equivalent TOPS and TDP in watts, with values sorted by TOPS per watt.Source: §4.2 · p. 7, Table 2 · p. 8

4.3 Model Training

• YOLO11 is used across experiments for compatibility with deployment targets, and pretrained YOLO11-Large models are fine-tuned for 100 epochs at 640^2 and 1024^2 pixel input resolutions.Unless otherwise stated, hyperparameters remain at their defaults.Source: §4.3 · p. 8

4.4 Model Deployment

• Models are exported to ONNX for broad CPU and GPU compatibility, OpenVINO is evaluated for CPU INT8 quantization, and Hailo deployments use the Dataflow Compiler with an optimized NMS block.Hailo-8 and Hailo-8L models are compiled separately using calibration images selected from the training set.Source: §4.4 · p. 9

4.5 Real-Time Optimization and Power Constraints

• Inference times are averaged over 500 randomized dummy-input frames per device–model combination after a 100-frame warm-up.Runtime measurements cover trained YOLO11 models from Nano to X-Large.Source: Runtime Measurements · p. 9
• On Raspberry Pi 5 with Hailo-8, total power and processed images are measured over time to derive average energy per inference, with idle draw recorded separately.The study treats energy per unit of useful work at a fixed camera-stream frame rate as the practical system-level measure.Source: Power Requirements · p. 9, §4.5 · p. 9

5.1 Detection Accuracy

• Hailo-8 and Hailo-8L achieve 58.7% and 56.2% mAP50–95, respectively; ARM CPU INT8 accuracy remained low despite export and calibration efforts.The authors suggest ARM limitations in accumulating 8-bit integer operations may explain the CPU result.Source: §5.1 · pp. 10–11, Table 3 · p. 10
• Detection accuracy is lowest in the Front scenario and highest in the Ground scenario, and is better under diffuse than mixed illumination; location effects are confounded with lighting.The scenario comparison uses the YOLO11-Large FP32 PyTorch model trained on Core-Manual; Fig. 8 reports image-averaged mAP50–95 deltas in percentage points.Source: §5.1 · p. 11, Figure 8 · p. 11
• At 1024^2 pixels, Ext-Auto reaches 50.6% and Full-Auto 52.6% mAP50–95; Core-Auto scores 42.7% at 640^2 and 49.5% at 1024^2, while the manual baseline consistently performs better than automatic-only training.Increasing input resolution benefits automatically annotated variants more than the manual baseline.Source: §5.1 · pp. 11–12, Table 4 · p. 12

5.2 Runtime Analysis

• At 640^2 pixels and FP32, Raspberry Pi 5 with Hailo-8 is second-fastest, 2.5–6.8× slower than the RTX 3090 but 37.5%–53.6% faster than the workstation CPU.Adding Hailo-8 yields a 29–45× speed-up over Raspberry Pi 5 CPU-only inference; ONNX is about 10% faster than PyTorch, and OpenVINO provides up to 7.5× speed-up from FP32 to INT8 on the AMD Ryzen 9 9950X3D.Source: Table 5 · p. 12, §5.2 · pp. 12–13

5.3 Power Usage

• At 640^2, the Hailo-optimized YOLO11-Small reaches 52.9% mAP50–95 and processes 825,390 images at 272.9 mJ per frame; the system sustains inference for over 7.5 hours on a 20,000 mAh power bank.Average power during inference is 8.29 W, compared with 3.4 W idle.Source: §5.3 · p. 13

6 Conclusion

• SAM 3 recovered 90% of manually identified beet instances with the broader prompt “vegetable,” achieving 46.1% mAP50–95 relative to manual annotations.The domain-specific prompt “sugar beet” did not yield reliable segmentation.Source: §6 · pp. 13–14
• At 640^2 pixels, manual-label YOLO11-Large reaches 56.9% mAP50–95 versus 50.6% for automatic labels; at 1024^2, the scores are 55.3% and 52.6%, respectively.Manual bounding-box annotation required approximately 25 hours, highlighting the effort saved by automated annotation at scale.Source: §6 · p. 14
• Deployment performance depends on platform-specific optimization: OpenVINO INT8 is fastest on x86-64 CPUs, while OpenVINO FP16 performs best on the Raspberry Pi 5 CPU.A Hailo-8 accelerator supports sustained, untethered operation using commodity USB power banks.Source: §6 · p. 14
• Future work should test scaling and robustness across lighting, occlusion, and conveyor speeds, and explore annotation filtering, self-training, and prompt ensembles.The outlook also proposes segmentation, tracking, and scene-scale estimation for additional operational metrics.Source: Outlook · p. 14

• The supplementary material adds details and analyses that complement the main paper’s benchmark description and quantitative results.Source: RootDown: Deploying and Benchmarking… · p. 18

A Extended Benchmark Statistics

• The extended Core-Manual statistics tabulate sample and recording-session totals for each state of the annotated parameters.Source: Table 6 · p. 18, A Extended Benchmark Statistics · p. 18

B Extended Quantitative Results

• Table 7 measures Scenario, Lighting, and Location effects on Core-Manual test samples using image-averaged mAP50-95 and deltas from 69.8%.This image-averaged metric is not directly comparable to the pooled evaluation in Table 3.Source: B Extended Quantitative Results · p. 18, Table 7 · p. 19
• Resolution-related performance differences are generally negligible, except for the OpenVINO INT8 model on Raspberry Pi 5.Table 8 covers YOLO11-Large detection performance at 1024² across deployment platforms, formats, and precisions.Source: B Extended Quantitative Results · p. 18, Table 8 · p. 19
• Tables 8 and 9 report YOLO11-Large detection performance and inference time, respectively, across deployment platforms, model formats, and numerical precisions at 1024².Hailo accelerators were evaluated exclusively as Raspberry Pi 5 HATs.Source: Table 8 · p. 19, Table 9 · p. 19