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

Gerardus Croonen; Andreas Trondl; Julia Simon; Daniel Steininger

cs.CVOct 23, 2026

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

Authors: Gerardus Croonen, Andreas Trondl, Julia Simon, Daniel Steininger

Organizations: AIT Austrian Institute of Technology, Giefinggasse 4, 1210 Vienna, Austria Center for Vision, Automation & Control

Abstract

Quality inspection of sugar beets after harvest and storage is critical to preserve sugar content and ensure optimal processing efficiency. The transfer of harvested beets to field clamps or trucks provides an opportunity to visually inspect individual beets automatically on loader conveyors. However, this is a challenging task as the large number of beets causes partial overlap and occlusion, complicating both annotation and detection. In this work, we propose an end-to-end image-based pipeline for detecting partially occluded sugar beets in loading scenarios as a first step toward automated quality inspection. We introduce a dedicated benchmark and present an automatic annotation strategy that leverages the SAM 3 foundation model to generate bounding boxes. We compare models trained on manually annotated and automatically generated labels. Additionally, we investigate deployment considerations by benchmarking multiple inference configurations, including execution on an edge device, under realistic operational constraints regarding precision, inference time and power consumption. Our results show that, despite substantial occlusion and inherent annotation ambiguity, foundation-model-based labels can substitute for a substantial share of manual annotation, trading labeling effort for data volume at a modest accuracy cost. Furthermore, we demonstrate that a Raspberry Pi 5 equipped with a Hailo accelerator enables high-performance, low-power inference, making it a practical solution for mobile outdoor deployment.

Figures & tables

Figure 1: Overview of the proposed pipeline and evaluation protocol, spanning data acquisition, annotation, training, evaluation and deployment on edge hardware. Multiple benchmark variants (pink) are defined using manually and automatically annotated data to assess both automatic annotation quality and deployment characteristics (orange, yellow and green: manually annotated training, validation and test splits; light and dark blue: automatically annotated training and validation splits).

Figure 2: Camera positions and viewpoints used in three scenarios: Front (A/B, top view), Side (C, side view) and Back (C, top view).

Figure 3: Relative distributions of annotated benchmark samples across acquisition scenarios, lighting conditions and recording locations.

Figure 4: Representative image samples illustrating the appearance variability across the benchmark scenarios Front , Side , Back and Ground .

Figure 5: Representative instance masks generated by SAM 3 using the text prompt vegetable , overlaid on corresponding sample images for each scenario.

	Annotation			Ratio			Images
Benchmark	Train	Val	Test	Train	Val	Test	Train	Val	Test	Total
Core - Manual	Man	Man	Man	50	20	30	156	62	94	312
Core - Auto	Auto	Auto	Man	50	20	30	156	62	94	312
Ext - Auto	Auto	Auto	Man	55	24	21	250	110	94	454
Full - Auto	Auto	Auto	Man	59	26	16	350	153	94	597

Table 1: Benchmark variants used for experimentation, all based on the same image split. To ensure fair comparison, the test-set images and annotations are identical across all variants and consist exclusively of manually annotated data.

Figure 6: Raspberry Pi 5 single-board computer with a Hailo-8 AI accelerator extension board mounted on top.

Platform	Type	Design	TOPS	TDP
Intel W-2125	CPU	x86-64	1.2	120
Raspberry Pi 5	CPU	ARM64	0.6	7
AMD 7430U	CPU	x86-64	1.5	15
AMD 9950X3D	CPU	x86-64	20.0	170
NVIDIA RTX 3090	GPU	CUDA	568.0	350
Hailo-8 on RPi 5	NPU	Custom	26.0	6

Table 2: Overview of the evaluated hardware platforms, showing INT8-equivalent TOPS (trillion operations per second) and TDP (thermal design power) in watts, sorted by TOPS per watt. Values shown in italics denote estimates derived from vendor-reported throughput figures.

Figure 7: Representative detection results on test images for a YOLO11-Large model with an input resolution of $640^{2}$ pixels deployed on a Raspberry Pi 5 using OpenVINO in FP32 and INT8 precision (green: correct detections, red: incorrect detections, blue: missed instances).

	PyTorch	ONNX		OpenVINO			Hailo
Platform	FP32	FP32	FP16	FP32	FP16	INT8	8	8L
Raspberry Pi 5	:	:	:	56.0	56.0	23.3	58.7	56.2
Intel W-2125	:	:	:	56.0	56.0	56.3	x	x
AMD 7430U	56.9	56.0	56.0	56.0	56.0	56.2	x	x
AMD 9950X3D	:	:	:	55.8	55.9	56.2	x	x
NVIDIA RTX 3090	:	:	:	x	x	x	x	x

Table 3: Detection performance (% mAP 50–95 ) of YOLO11-Large models at an input resolution of

640^{2}

pixels trained on the Core-Manual benchmark variant across different deployment platforms, model formats and numerical precisions. An x indicates unsupported combinations, a colon denotes identical values across a column. Hailo accelerators were evaluated exclusively as Raspberry Pi 5 HATs.

Figure 8: Impact of meta-parameters on detection performance for a YOLO11-Large FP32 PyTorch model trained on the Core-Manual benchmark variant. Deltas are computed from image-averaged mAP 50–95 and given in percentage points.

Benchmark	$640^{2}$	$1024^{2}$
Core-Auto	42.7	49.5
Ext-Auto	48.0	49.7
Full-Auto	50.6	52.6
Core-Manual	56.9	55.3

Table 4: Detection performance (% mAP 50–95 ) of YOLO11-Large models across all benchmark variants at input resolutions of

640^{2}

and

1024^{2}

pixels.

Platform	N	S	M	L	X
Raspberry Pi 5	287	610	1,563	1,961	3,704
Intel W-2125	95	149	250	435	667
AMD 7430U	61	139	357	476	909
AMD 9950X3D	16	36	74	91	179
NVIDIA RTX 3090	4	5	6	11	20
Hailo-8L on RPi 5	13	26	44	70	116

Table 5: Inference time (ms) across compute platforms and YOLO11 model sizes ( N ano, S mall, M edium, L arge and X -Large) at FP32 precision (where applicable) and an input resolution of

640^{2}

. Best results are marked in bold.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Parameter	State	Rec	Samples
Scenario	Front	18	64
	Side	2	10
	Back	35	130
	Ground	1	108
Lighting	Sunny	21	78
	Mixed	7	27

Appendix

Table 6: Detailed benchmark statistics for all annotated parameters of the Core-Manual benchmark, showing the total numbers of Rec ording sessions and Samples for each state.

Parameter	State	Score	Delta
Scenario	Front	60.0	-9.8
	Side	69.9	0.0
	Back	68.6	-1.2
	Ground	78.8	9.0
Lighting	Sunny	68.3	-1.5
	Mixed	63.9	-5.9

Appendix

Table 7: Impact of the meta-parameters Scenario , Lighting and Location on model performance, measured on the test samples of the Core-Manual benchmark. Scores are image-averaged mAP 50-95 (%), i.e. mAP is computed per image and averaged over the subset, and Delta gives the difference to the overall image-averaged score of 69.8% in percentage points. This metric is not directly comparable to the pooled evaluation in Tab. 3, which ranks all detections of the test set jointly.

	PyTorch	ONNX		OpenVINO			Hailo
Platform	FP32	FP32	FP16	FP32	FP16	INT8	8	8L
Raspberry Pi 5	:	:	54.2	:	54.3	16.3	55.7	55.1
Intel W-2125	:	:	54.3	:	54.3	55.1	x	x
AMD 7430U	55.3	54.3	54.3	54.3	54.3	55.2	x	x
AMD 9950X3D	:	:	54.3	:	54.1	55.4	x	x
NVIDIA RTX 3090	:	:	54.3	x	x	x	x	x

Appendix

Table 8: Detection performance (% mAP 50–95 ) of YOLO11-Large models at an input resolution of

1024^{2}

pixels across different deployment platforms, model formats (PyTorch, ONNX and OpenVINO) and numerical precisions (FP32, FP16 and INT8). An x indicates unsupported combinations, a colon denotes identical values across a column. Hailo accelerators were evaluated exclusively as Raspberry Pi 5 HATs.

Platform	FP32	FP32	FP16	FP32	FP16	INT8	8	8L
	PyTorch	ONNX		OpenVINO			Hailo
Raspberry Pi 5	1,961	1,299	1,299	568	529	1,136	47	70
Intel W-2125	435	345	238	172	185	114	x	x
AMD 7430U	476	345	345	250	244	118	x	x
AMD 9950X3D	91	128	128	19	19	12	x	x
NVIDIA RTX 3090	11	9	9	x	x	x	x	x

Appendix

Table 9: Inference time (ms) of YOLO11-Large models at an input resolution of

1024^{2}

pixels across different deployment platforms, model formats (PyTorch, ONNX and OpenVINO) and numerical precisions (FP32, FP16 and INT8). An x indicates unsupported combinations. Hailo accelerators were evaluated exclusively as Raspberry Pi 5 HATs.

Object Detection

Object detection locates and classifies objects in images or video, balancing localization accuracy, robustness, and inference speed.

See all 624 papers

SAM3-Assisted Training of Lightweight YOLO Models for Precision Pig Farming
May 25, 2026Marcos Vinicius Mendes Faria, Thiago Borges Pereira, Isabella C. F. S. Condotta +2Edge Inference Pseudo-Labeling
Coordenadoria de Inform´atica, Instituto Federal do Esp´ırito Santo (IFES), Serra, ES · Department of Animal Sciences, University of Illinois at Urbana-Champaign, USA
A Comprehensive Evaluation of Deep Learning Object Detection Models on Heterogeneous Edge Devices
Sep 25, 2024Daghash K. Alqahtani, Muhammad Aamir Cheema, Maria A. Rodriguez +1Energy-Efficient ML Edge Inference
DisNet Lab, School of Computing and Information Systems, The University of Melbourne, Melbourne, VIC, Australia · Monash University, Melbourne, VIC, Australia
TinyBayes: Closed-Form Bayesian Inference via Jacobi Prior for Real-Time Image Classification on Edge Devices
May 7, 2026Shouvik Sardar, Sourish DasPlant Disease Classification Edge Inference
Data Science Group Chennai Mathematical Institute

May 25, 2026J/K move · Enter open · S save

SAM3-Assisted Training of Lightweight YOLO Models for Precision Pig Farming

Marcos Vinicius Mendes Faria, Thiago Borges Pereira, Isabella C. F. S. Condotta, Thiago Meireles Paixão+1

Coordenadoria de Inform´atica, Instituto Federal do Esp´ırito Santo (IFES), Serra, ES · Department of Animal Sciences, University of Illinois at Urbana-Champaign, USA

Edge Inference Pseudo-Labeling Object Detection Real-Time Object Detection Knowledge Distillation Precision Livestock Farming

Open paper AskPDF arXiv

Deep learning-based object detection has revolutionized Precision Livestock Farming (PLF), yet a critical barrier remains: high-performance Foundation Models (such as SAM 3) are too computationally intensive for edge deployment, while lightweight models (like YOLO) require prohibitive manual annotation efforts. This work proposes a fully automated knowledge distillation pipeline that leverages the Segment Anything Model 3 (SAM 3) to generate zero-shot pseudo-labels for training efficient YOLOv8 detectors. By treating SAM 3 as an offline auto-annotator, we eliminate the manual labeling bottleneck, producing models capable of real-time inference on resource-constrained hardware. We systematically evaluate this approach on the PigLife dataset, comparing SAM 3-supervised models against human-annotated baselines. Results demonstrate that a SAM 3-trained YOLOv8m achieves a mean Average Precision (mAP) of 79.4% without human intervention, while reducing inference latency by approximately 200

\times

compared to the teacher model. Furthermore, stratified analysis reveals that in low-occlusion scenarios, the automated pipeline achieves detection rates comparable to human benchmarks (

AP_{50} > 99\%

). These findings indicate that foundation models can serve as effective, zero-annotation-cost supervisors, enabling scalable edge computing solutions for smart agriculture.

Abstract

Object Detection

Explore similar work

SAM3-Assisted Training of Lightweight YOLO Models for Precision Pig Farming

A Comprehensive Evaluation of Deep Learning Object Detection Models on Heterogeneous Edge Devices

TinyBayes: Closed-Form Bayesian Inference via Jacobi Prior for Real-Time Image Classification on Edge Devices