A Comparative Study of Modern Object Detectors for Robust Apple Detection in Orchard Imagery

Mohammed Asad; Ajai Kumar Gautam; Priyanshu Dhiman; Rishi Raj Prajapati

cs.CVApr 11, 2026

A Comparative Study of Modern Object Detectors for Robust Apple Detection in Orchard Imagery

Authors: Mohammed Asad, Ajai Kumar Gautam, Priyanshu Dhiman, Rishi Raj Prajapati

Organizations: Dept. of Electronics and Communications Engineering, Delhi Technological University, New Delhi, India

Abstract

Accurate apple detection in orchard images is important for yield prediction, fruit counting, robotic harvesting, and crop monitoring. However, changing illumination, leaf clutter, dense fruit clusters, and partial occlusion make detection difficult. To provide a fair and reproducible comparison, this study establishes a controlled benchmark for single-class apple detection on the public AppleBBCH81 dataset using one deterministic train, validation, and test split and a unified evaluation protocol across six representative detectors: YOLOv10n, YOLO11n, RT-DETR-L, Faster R-CNN (ResNet50-FPN), FCOS (ResNet50-FPN), and SSDLite320 (MobileNetV3-Large). Performance is evaluated primarily using COCO-style [email protected] and [email protected]:0.95, and threshold-dependent behavior is further analyzed using precision-recall curves and fixed-threshold precision, recall, and F1-score at IoU = 0.5. On the validation split, YOLO11n achieves the best strict localization performance with [email protected]:0.95 = 0.6065 and [email protected] = 0.9620, followed closely by RT-DETR-L and YOLOv10n. At a fixed operating point with confidence >= 0.05, YOLOv10n attains the highest F1-score, whereas RT-DETR-L achieves very high recall but low precision because of many false positives at low confidence. These findings show that detector selection for orchard deployment should be guided not only by localization-aware accuracy but also by threshold robustness and the requirements of the downstream task.

Figures & tables

Study	Main focus	Strength	Limitation relative to this study
DeepFruits [ 10 ]	Deep learning-based fruit detection in orchards	Demonstrated the feasibility of robust learning-based fruit detection in field imagery	Older detection setting; not a controlled benchmark across modern detector families
Bargoti and Underwood [ 11 ]	Fruit detection under orchard clutter and occlusion	Highlighted the importance of real orchard conditions, dataset quality, and occlusion handling	Did not provide a unified cross-family benchmark under one evaluation protocol
MinneApple [ 12 ]	Apple detection and segmentation dataset	Enabled standardized evaluation on a public apple dataset	Focused on dataset creation; not a controlled comparison of multiple modern detector families on AppleBBCH81
Hani et al. [ 25 ]	Fruit detection versus counting performance	Showed that detection quality and counting quality are related but not identical	Focused on detection and counting objectives rather than detector-family benchmarking under one shared protocol
BFPNet [ 15 ]	Small apple detection in complex orchard scenes	Improved small-object recall using a feature-pyramid design	Specialized architecture study rather than a broad detector-family comparison
Ma et al. [ 16 ]	Improved YOLO-based detection and counting of small apples	Strong task-specific optimization for challenging environments	Single-architecture focus; limited cross-family comparison

TABLE I: Summary of representative prior work on fruit and apple detection and their limitations relative to this study.

Fig. 1: Proposed workflow for apple detection on AppleBBCH81.

Model	Framework	Initialization	Input size	Batch	Epochs	Optimizer	Learning rate
YOLOv10n	Ultralytics	COCO pretrained	640	16	50	auto	0.01
YOLO11n	Ultralytics	COCO pretrained	640	16	50	auto	0.01
RT-DETR-L	Ultralytics	COCO pretrained	640	16	50	auto	0.01
Faster R-CNN (R50-FPN)	Torchvision	COCO pretrained	800 to 1333	4	10	AdamW	$1\times 10^{-4}$
FCOS (R50-FPN)	Torchvision	COCO pretrained	800 to 1333	4	10	AdamW	$1\times 10^{-4}$
SSDLite320 (MobileNetV3-Large)	Torchvision	COCO pretrained	320	4	10	AdamW	$1\times 10^{-4}$

TABLE II: Summary of model training configurations used in the benchmark.

Fig. 2: Model taxonomy of the evaluated detectors.

Family	Model	[email protected]:0.95	[email protected]
Ultralytics	YOLO11n	0.6065	0.9620
Ultralytics	RT-DETR-L	0.6012	0.9506
Ultralytics	YOLOv10n	0.5941	0.9501
Torchvision	Faster R-CNN (R50-FPN)	0.5278	0.9091
Torchvision	FCOS (R50-FPN)	0.4926	0.8993
Torchvision	SSDLite320 (MobileNetV3-Large)	0.2125	0.5400

TABLE III: COCO-style validation results for single-class apple detection.

Model	TP	FP	FN	PR AUC	Precision	95% CI	Recall	95% CI	F1
YOLO11n	1465	920	46	0.958	0.614256	[0.594551, 0.633593]	0.969557	[0.959632, 0.977099]	0.752053
RT-DETR-L	1466	5061	45	0.950	0.224605	[0.214645, 0.234890]	0.970218	[0.960383, 0.977669]	0.364767
YOLOv10n	1445	745	66	0.947	0.659817	[0.639710, 0.679365]	0.956320	[0.944806, 0.965520]	0.780870
Faster R-CNN (R50-FPN)	1391	717	120	0.900	0.659867	[0.639368, 0.679784]	0.920582	[0.905858, 0.933174]	0.768721
FCOS (R50-FPN)	1417	1289	94	0.904	0.523651	[0.504813, 0.542422]	0.937790	[0.924465, 0.948893]	0.672042
SSDLite320 (MobileNetV3-Large)	1323	30213	188	0.542	0.041952	[0.039795, 0.044221]	0.875579	[0.857978, 0.891275]	0.080068

TABLE IV: Validation diagnostics at IoU

=0.5

and confidence

\geq 0.05

using one-to-one greedy matching.

Fig. 3: Precision-recall curves for all models on the validation split at IoU $=0.5$ .

Fig. 4: Object-level confusion summary for YOLO11n on the validation split at IoU $=0.5$ and confidence $\geq 0.05$ (TP = 1465, FP = 920, FN = 46; TN is not defined for object detection).

Explore similar work

Optimizing YOLO27, YOLO26, YOLO11, and YOLOv8 for Fine-Grained Small-Object Detection and Segmentation in Complex Orchard Environments
Aug 23, 2026Ranjan Sapkota, Manoj KarkeeSmall Object Detection Instance Segmentation
Cornell University, AgRobotics Lab, 134 Riley-Robb Hall, Ithaca, NY, 14853, USA
A Lightweight Multimodal Vision-Language Framework for Early-Stage Anatomical Green Fruit Classification in Commercial Orchards
Aug 23, 2026Ranjan Sapkota, William Bu, Chen Chen +2Agricultural Image Analysis Fine-Grained Image Classification
Cornell University, Department of Environmental and Biological Engineering, USA · Department of Computer Science, University of Central Florida, USA · Institute of Artificial Intelligence (IAI) & Department of Computer Science, University of Central Florida, USA +1
Towards Practical Precision Agriculture: Real-Time Fruit Detection and Video Analytics on Embedded Edge Hardware
Sep 11, 2026Ivica Dimitrovski, Vlatko Spasev, Ivan Kitanovski +2Agricultural Image Analysis Real-Time Object Detection
Faculty of Computer Science and Engineering, University Ss Cyril and Methodius, ul. Rudzer Boshkovikj 16, P.O. 393, Skopje, 100, North Macedonia · Faculty of Agricultural Sciences and Food, University Ss Cyril and Methodius, 16-ta Makedonska brigada br. 3, Skopje, 100, North Macedonia

Aug 23, 2026J/K move · Enter open · S save

Optimizing YOLO27, YOLO26, YOLO11, and YOLOv8 for Fine-Grained Small-Object Detection and Segmentation in Complex Orchard Environments

Ranjan Sapkota, Manoj Karkee

Cornell University, AgRobotics Lab, 134 Riley-Robb Hall, Ithaca, NY, 14853, USA

Small Object Detection Instance Segmentation Fine-Grained Visual Perception You Only Look Once Agricultural Robotics Robotic Perception

Open paper AskPDF arXiv

This study presents an architectural and experimental cross-generation analysis of Ultralytics YOLO27 (YOLOv27), YOLO26 (YOLOv26), YOLO11 (YOLOv11), and YOLOv8 for fine-grained robotic perception in complex orchard environments. Fine-grained detection and instance segmentation of early-stage fruit anatomy remain challenging in complex orchard environments because of limited pixel footprints, green-on-green similarity, occlusion, and substantial scale variation. Because YOLO27 has been announced but its public implementation and trainable segmentation models are not yet available, the present study provides an architectural analysis of YOLO27, while controlled experiments benchmark YOLOv8, YOLO11, and YOLO26; YOLO27 experiments will be incorporated following public model availability. Five model scales-nano (n), small (s), medium (m), large (l), and extra-large (x)-were evaluated for fruitlet, calyx, and peduncle detection and segmentation using conventional 640 x 640 and small-object-focused 960 x 960 configurations, yielding 30 experiments. YOLO11s-960 achieved the highest observed mask mAP@50:95(0.402) and box mAP@50:95(0.426), whereas YOLO26s-960 achieved comparable values of 0.397 and 0.425 with only 10.37~~M parameters and 34.1~~GFLOPs. Peduncle remained the most challenging class, and increasing model capacity did not consistently improve accuracy. Overall, compact-to-moderate YOLO models combined with small-object-focused training provided favorable accuracy-efficiency trade-offs, establishing a reproducible benchmark for fine-grained agricultural robotic perception. Code, trained models, and experimental configurations are publicly available, and will be updated through our Github Link: https://github.com/rnjnspkt/Optimizing-and-Comparing-Ultralytics-YOLOv26-YOLOv11-and-YOLOv8-for-Small-Object-Detection-and-Seg

Abstract

Explore similar work

Optimizing YOLO27, YOLO26, YOLO11, and YOLOv8 for Fine-Grained Small-Object Detection and Segmentation in Complex Orchard Environments

A Lightweight Multimodal Vision-Language Framework for Early-Stage Anatomical Green Fruit Classification in Commercial Orchards

Towards Practical Precision Agriculture: Real-Time Fruit Detection and Video Analytics on Embedded Edge Hardware