cs.CVAug 23, 2026

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

Authors: Ranjan Sapkota, Manoj Karkee

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

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

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 10.37M parameters and 34.1GFLOPs. 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

Figures & tables

Figure 1: Accuracy–computational-complexity comparison of YOLO27, YOLO26, YOLO11, YOLOv10, YOLOv9, YOLOv8, and YOLOv7 detection models on COCO. The vertical axis reports mAP 50:95val{}_{50:95}^{val} , while the horizontal axis represents floating-point operations at approximately 640-pixel input resolution. YOLO27 and YOLO26 are emphasized to illustrate the transition from deployment-oriented CNN/end-to-end detection toward scale-adaptive architectures, including query-based YOLO27m/l detection. FLOPs provide a hardware-independent approximation of arithmetic complexity but do not capture memory movement, operator fusion, attention-kernel efficiency, or accelerator utilization. YOLO27 results are preliminary.
Figure 2: Scale-adaptive dual architecture of YOLO27. Compact YOLO27n/s detectors employ a streamlined CNN-based pathway with strengthened high-resolution representations, FPN/PAN feature fusion, fine- and coarse-scale prediction, foreground-alignment supervision, and either one-to-many NMS-based or one-to-one NMS-free detection. In contrast, YOLO27m/l employ query-based detection with learned object queries and transformer decoding. YOLO27m retains a CNN backbone, whereas YOLO27l incorporates an UltraViT backbone. Dashed connections denote training-specific supervision rather than the principal inference pathway.
Figure 3: Representative examples of complex green fruit orchard scenes used in this study. (a) Full-view image of a commercial apple orchard during the early green fruit (fruitlet) development stage, where immature fruitlets are visually camouflaged within dense green canopy foliage. (b) Enlarged canopy region highlighting a dense cluster of approximately ten green fruitlets (yellow dashed circle), illustrating high object proximity, clustering, and background clutter. (c) Fine-grained anatomical annotation of clustered fruitlets into calyx, fruit main body, and peduncle, each representing a small and low-contrast target. (d) Workers performing manual fruitlet thinning in the field, a labor-intensive and costly but essential tree-fruit production operation increasingly constrained by worker shortages.
Figure 4: Experimental workflow for cross-generation YOLO optimization. Orchard data preparation is followed by training and validation of YOLOv8, YOLO11, and YOLO26 variants under conventional and small-object-focused configurations, with subsequent comparative assessment of detection, segmentation, computational efficiency, and deployment suitability.
Figure 5: Overview of the study site and data collection environment for small-object detection and segmentation; (a) Geographical map of Washington State highlighting the location where the data were collected (Prosser, WA); b) Commercial apple orchard used in this study, illustrating the real-world field conditions during image acquisition.
Figure 6: Detailed architecture of YOLO27n/s. The architecture incorporates a streamlined CNN backbone, widened high-resolution features, FPN/PAN feature fusion, and dual-scale prediction using fine ( PfP_{f} ) and coarse ( PcP_{c} ) feature maps. Training combines one-to-many supervision with a foreground-alignment branch, whereas deployment uses one-to-one predictions for efficient NMS-free inference.
Figure 7: Architectural overview of the model families and deployment concept considered in this study: (a) YOLO26, (b) YOLO11, and (c) YOLOv8
ComponentSpecification
Processor (CPU)AMD Ryzen Threadripper 7960
CPU cores / threads24 cores / 48 threads
System memory256 GiB RAM
Graphics processorNVIDIA RTX 5000 Ada Generation
Reported GPU memory32,237 MiB CUDA memory ( ∼\sim 33.8 GB device-reported capacity)
GPU deviceCUDA device 0
Table 1: Hardware and software environment used for cross-generation YOLO training and evaluation.
Training ComponentConventional ConfigurationSmall-Object ConfigurationImplementation Detail
YOLO familiesYOLOv8 / YOLO11 / YOLO26YOLOv8 / YOLO11 / YOLO26Instance-segmentation models
Model scalesn, s, m, l, xn, s, m, l, xFive scales per generation
Total architectures1515Three generations ×\times five scales
Model initializationPretrained segmentation weightsPretrained segmentation weightspretrained=True
Input image size640×640640\times 640960×960960\times 960Square model-training input
Maximum training epochs200200Nominal principal training schedule
Table 2: Training and augmentation configurations used for the cross-generation YOLO experiments. The conventional configuration used 640×640640\times 640 -pixel inputs, whereas the small-object-focused configuration used 960×960960\times 960 -pixel inputs with additional augmentation controls. Parameters identified as framework-resolved were obtained from the recorded Ultralytics training configuration.
Figure 8: Qualitative comparison of cross-generation YOLO detection and instance segmentation under conventional and small-object-focused training configurations. The leftmost panel shows the representative original orchard image, followed by prediction and corresponding heatmap visualizations obtained using YOLO26, YOLO11, and YOLOv8 at 640×640640\times 640 and 960×960960\times 960 input resolutions. The examples illustrate differences in the localization and segmentation of calyx, fruitlet, and peduncle structures under challenging green-on-green conditions. Overall, the 960×960960\times 960 configurations recover additional small fruitlet instances in several regions, although model-specific differences remain evident, particularly for the substantially smaller and less represented peduncle class.
ModelResolutionPRBox mAP 50Box mAP 50:95Mask mAP 50Mask mAP 50:95
YOLO26n-seg6400.5680.7040.6610.3630.6440.335
YOLO26n-seg9600.6610.6970.6820.4140.6760.384
YOLO26s-seg6400.5650.6460.6270.3770.6070.341
YOLO26s-seg9600.6180.7510.6870.4250.6860.397
YOLO26m-seg6400.6480.6410.6540.4030.6290.361
YOLO26m-seg9600.5880.7280.6650.4190.6650.395
Table 3: Aggregate detection and instance-segmentation performance of YOLO26 segmentation models under 640 ×\times 640 and 960 ×\times 960 input resolutions. Results correspond to the aggregate “All” category across calyx, fruitlet, and peduncle classes. Precision (P) and recall (R) correspond to bounding-box detection, while box and mask mAP are reported at IoU = 0.50 and over IoU = 0.50–0.95. Bold values indicate the highest value obtained for each metric across all evaluated YOLO26 configurations.
ModelΔ\Delta PΔ\Delta RΔ\Delta Box mAP 50Δ\Delta Box mAP 50:95Δ\Delta Mask mAP 50Δ\Delta Mask mAP 50:95
YOLO26n-seg+0.093-0.007+0.021+0.051+0.032+0.049
YOLO26s-seg+0.053+0.105+0.060+0.048+0.079+0.056
YOLO26m-seg-0.060+0.087+0.011+0.016+0.036+0.034
YOLO26l-seg-0.037-0.029-0.028+0.006+0.011+0.020
YOLO26x-seg+0.111-0.134-0.009+0.009-0.017-0.003
Table 4: Absolute change in aggregate YOLO26 performance when increasing input resolution from 640 ×\times 640 to 960 ×\times 960 pixels. Changes were calculated as Δ=Metric960−Metric640\Delta=\mathrm{Metric}_{960}-\mathrm{Metric}_{640} .
ModelResolutionLayersParametersGFLOPsPreprocessInferencePostprocess
     (ms)(ms)(ms)
YOLO26n-seg6401392,689,4699.00.11.90.3
YOLO26n-seg9601392,689,4699.00.21.80.6
YOLO26s-seg64013910,366,50134.10.31.80.4
YOLO26s-seg96013910,366,50134.10.23.00.8
YOLO26m-seg64014923,509,781121.20.13.00.4
Table 5: Architectural complexity and validation-time processing speed of YOLO26 segmentation models under 640 ×\times 640 and 960 ×\times 960 input resolutions. Lower preprocessing, inference, and postprocessing times indicate greater computational efficiency, whereas higher parameter counts and GFLOPs indicate greater model capacity and computational demand.
Figure 9: Training–validation precision–recall (PR) curves for YOLO26 instance-segmentation models trained using the conventional 640×640640\times 640 -pixel input-resolution configuration: (a,b) YOLO26n-seg, (c,d) YOLO26s-seg, (e,f) YOLO26m-seg, (g,h) YOLO26l-seg, and (i,j) YOLO26x-seg. Panels (a,c,e,g,i) present bounding-box PR curves, whereas panels (b,d,f,h,j) present instance-mask PR curves.
Figure 10: Training–validation precision–recall (PR) curves for YOLO26 instance-segmentation models trained using the small-object-focused 960×960960\times 960 -pixel input-resolution configuration: (a,b) YOLO26n-seg, (c,d) YOLO26s-seg, (e,f) YOLO26m-seg, (g,h) YOLO26l-seg, and (i,j) YOLO26x-seg. Panels (a,c,e,g,i) present bounding-box PR curves, whereas panels (b,d,f,h,j) present instance-mask PR curves.
ModelResolutionPRBox mAP 50Box mAP 50:95Mask mAP 50Mask mAP 50:95
YOLO11n-seg6400.5460.6900.6370.3570.6220.318
YOLO11n-seg9600.5950.7050.6660.3960.6690.367
YOLO11s-seg6400.5530.6680.6230.3640.6210.334
YOLO11s-seg9600.6680.6410.6940.4260.6920.402
YOLO11m-seg6400.6120.6990.6800.4040.6780.361
YOLO11m-seg9600.6510.6420.6620.4140.6610.382
Table 6: Aggregate detection and instance-segmentation performance of YOLO11 segmentation models under 640 ×\times 640 and 960 ×\times 960 input resolutions. Results correspond to the aggregate “All” category across calyx, fruitlet, and peduncle classes. Precision (P) and recall (R) correspond to bounding-box detection, while box and mask mAP are reported at IoU = 0.50 and over IoU = 0.50–0.95. Bold values indicate the highest value obtained for each metric across all YOLO11 configurations.
ModelΔ\Delta PΔ\Delta RΔ\Delta Box mAP 50Δ\Delta Box mAP 50:95Δ\Delta Mask mAP 50Δ\Delta Mask mAP 50:95
YOLO11n-seg+0.049+0.015+0.029+0.039+0.047+0.049
YOLO11s-seg+0.115-0.027+0.071+0.062+0.071+0.068
YOLO11m-seg+0.039-0.057-0.018+0.010-0.017+0.021
YOLO11l-seg-0.079+0.077+0.007+0.025+0.011+0.041
YOLO11x-seg-0.055+0.075+0.020+0.033+0.024+0.048
Table 7: Absolute change in aggregate YOLO11 performance when increasing input resolution from 640 ×\times 640 to 960 ×\times 960 pixels. Changes were calculated as Δ=Metric960−Metric640\Delta=\mathrm{Metric}_{960}-\mathrm{Metric}_{640} . Positive values indicate improvement under the higher-resolution configuration.
ModelResolutionLayersParametersGFLOPsPreprocessInferencePostprocess
     (ms)(ms)(ms)
YOLO11n-seg6401142,835,1539.60.81.02.4
YOLO11n-seg9601142,835,1539.60.52.81.9
YOLO11s-seg64011410,067,97732.80.82.33.5
YOLO11s-seg96011410,067,97732.80.33.81.6
YOLO11m-seg64013922,337,625112.90.83.82.4
Table 8: Architectural complexity and validation-time processing speed of YOLO11 segmentation models under 640 ×\times 640 and 960 ×\times 960 input resolutions. Lower preprocessing, inference, and postprocessing times indicate greater computational efficiency, whereas larger parameter counts and GFLOPs indicate greater architectural capacity and computational demand.
Figure 11: Training–validation precision–recall (PR) curves for YOLO11 instance-segmentation models trained using the conventional 640×640640\times 640 -pixel input-resolution configuration: (a,b) YOLO11n-seg, (c,d) YOLO11s-seg, (e,f) YOLO11m-seg, (g,h) YOLO11l-seg, and (i,j) YOLO11x-seg. Panels (a,c,e,g,i) present bounding-box PR curves, whereas panels (b,d,f,h,j) present instance-mask PR curves.
Figure 12: Training–validation precision–recall (PR) curves for YOLO11 instance-segmentation models trained using the small-object-focused 960×960960\times 960 -pixel input-resolution configuration: (a,b) YOLO11n-seg, (c,d) YOLO11s-seg, (e,f) YOLO11m-seg, (g,h) YOLO11l-seg, and (i,j) YOLO11x-seg. Panels (a,c,e,g,i) present bounding-box PR curves, whereas panels (b,d,f,h,j) present instance-mask PR curves.
ModelResolutionPRBox mAP 50Box mAP 50:95Mask mAP 50Mask mAP 50:95
YOLOv8n-seg6400.6280.7300.6600.3560.6490.325
YOLOv8n-seg9600.6190.6280.6460.3900.6520.371
YOLOv8s-seg6400.6560.6590.6660.3900.6480.356
YOLOv8s-seg9600.6050.7740.6920.4190.6900.390
YOLOv8m-seg6400.6280.7350.6260.3760.6090.337
YOLOv8m-seg9600.6400.6360.6600.3950.6540.371
Table 9: Aggregate detection and instance-segmentation performance of YOLOv8 segmentation models under 640 ×\times 640 and 960 ×\times 960 input resolutions. Results correspond to the aggregate “All” category across calyx, fruitlet, and peduncle classes. Precision (P) and recall (R) correspond to bounding-box detection. Bold values indicate the highest value obtained for each metric across all YOLOv8 configurations.
ModelΔ\Delta PΔ\Delta RΔ\Delta Box mAP 50Δ\Delta Box mAP 50:95Δ\Delta Mask mAP 50Δ\Delta Mask mAP 50:95
YOLOv8n-seg-0.009-0.102-0.014+0.034+0.003+0.046
YOLOv8s-seg-0.051+0.115+0.026+0.029+0.042+0.034
YOLOv8m-seg+0.012-0.099+0.034+0.019+0.045+0.034
YOLOv8l-seg+0.100-0.061+0.033+0.034+0.039+0.050
YOLOv8x-seg+0.039+0.059+0.023+0.039+0.032+0.045
Table 10: Absolute change in aggregate YOLOv8 performance when increasing input resolution from 640 ×\times 640 to 960 ×\times 960 pixels. Changes were calculated as Δ=Metric960−Metric640\Delta=\mathrm{Metric}_{960}-\mathrm{Metric}_{640} .
ModelResolutionLayersParametersGFLOPsPreprocessInferencePostprocess
     (ms)(ms)(ms)
YOLOv8n-seg640863,258,64911.30.41.22.2
YOLOv8n-seg960863,258,64911.30.21.72.3
YOLOv8s-seg6408611,780,76139.90.92.80.9
YOLOv8s-seg9608611,780,76139.90.42.40.9
YOLOv8m-seg64010627,224,121104.31.45.41.2
Table 11: Architectural complexity and validation-time processing speed of YOLOv8 segmentation models under 640 ×\times 640 and 960 ×\times 960 input resolutions. Lower processing times represent greater computational efficiency, whereas higher parameter counts and GFLOPs represent greater model capacity and computational demand.
Figure 13: Training–validation precision–recall (PR) curves for YOLOv8 instance-segmentation models trained using the conventional 640×640640\times 640 -pixel input-resolution configuration: (a,b) YOLOv8n-seg, (c,d) YOLOv8s-seg, (e,f) YOLOv8m-seg, (g,h) YOLOv8l-seg, and (i,j) YOLOv8x-seg. Panels (a,c,e,g,i) present bounding-box PR curves, whereas panels (b,d,f,h,j) present instance-mask PR curves.
Figure 14: Training–validation precision–recall (PR) curves for YOLOv8 instance-segmentation models trained using the small-object-focused 960×960960\times 960 -pixel input-resolution configuration: (a,b) YOLOv8n-seg, (c,d) YOLOv8s-seg, (e,f) YOLOv8m-seg, (g,h) YOLOv8l-seg, and (i,j) YOLOv8x-seg. Panels (a,c,e,g,i) present bounding-box PR curves, whereas panels (b,d,f,h,j) present instance-mask PR curves.
Figure 15: Qualitative comparison of YOLO26 instance-segmentation results at 640×640640\times 640 and 960×960960\times 960 training resolutions. Red dotted circles highlight additional calyx and fruitlet instances recovered by the higher-resolution configuration, whereas the peduncle remains undetected at both resolutions. Corresponding heatmaps illustrate the enhanced spatial response and anatomical-feature representation achieved with 960×960960\times 960 training.
ModelResolutionBounding BoxMask
  PRmAP 50mAP 50:95PRmAP 50mAP 50:95
YOLO26n-seg6400.6430.8410.7910.4790.6430.8410.7950.476
YOLO26n-seg9600.7550.7420.8030.5410.7540.7590.8110.519
YOLO26s-seg6400.7030.8010.8060.5300.6940.8020.8010.509
YOLO26s-seg9600.7150.8260.8030.5540.7380.7770.8080.534
YOLO26m-seg6400.7270.7960.8130.5580.7300.7990.8180.525
Table 12: Class-wise detection and instance-segmentation performance for the fruitlet class using YOLO26 models under 640 ×\times 640 and 960 ×\times 960 input resolutions. Bold values indicate the highest value for each metric.
ModelResolutionBounding BoxMask
  PRmAP 50mAP 50:95PRmAP 50mAP 50:95
YOLO26n-seg6400.6250.8130.7260.3600.6010.7810.6980.328
YOLO26n-seg9600.7460.7710.7560.3970.7170.7550.7370.371
YOLO26s-seg6400.6180.7710.7360.4010.6000.7550.7180.352
YOLO26s-seg9600.6640.7920.7100.3970.7040.7570.7070.366
YOLO26m-seg6400.6720.7420.6980.3750.6660.7360.6760.337
Table 13: Class-wise detection and instance-segmentation performance for the calyx class using YOLO26 models under 640 ×\times 640 and 960 ×\times 960 input resolutions. Bold values indicate the highest value for each metric.
ModelResolutionBounding BoxMask
  PRmAP 50mAP 50:95PRmAP 50mAP 50:95
YOLO26n-seg6400.4340.4570.4680.2500.4150.4380.4410.201
YOLO26n-seg9600.4830.5770.4870.3030.4780.5770.4800.263
YOLO26s-seg6400.3750.3650.3390.2000.3330.3370.3040.163
YOLO26s-seg9600.4770.6350.5490.3250.5860.5580.5420.290
YOLO26m-seg6400.5450.3850.4500.2780.4900.3460.3940.220
Table 14: Class-wise detection and instance-segmentation performance for the peduncle class using YOLO26 models under 640 ×\times 640 and 960 ×\times 960 input resolutions. Bold values indicate the highest value for each metric.
ClassΔ\Delta Box PΔ\Delta Box RΔ\Delta Box mAP 50Δ\Delta Box mAP 50:95Δ\Delta Mask PΔ\Delta Mask RΔ\Delta Mask mAP 50Δ\Delta Mask mAP 50:95
Fruitlet+0.019-0.020-0.005+0.019+0.019-0.024-0.001+0.022
Calyx+0.023+0.006-0.001+0.019+0.027+0.008+0.008+0.027
Peduncle+0.055+0.026+0.039+0.040+0.106+0.069+0.077+0.044
Table 15: Mean absolute change in class-wise YOLO26 performance when increasing input resolution from 640 ×\times 640 to 960 ×\times 960 pixels across all five model scales.
Figure 16: Qualitative comparison of YOLO11 segmentation at 640×640640\times 640 and 960×960960\times 960 training resolutions. Solid red circles highlight peduncles successfully detected and segmented only by the 960×960960\times 960 configuration, demonstrating improved fine-scale anatomical perception while both configurations effectively identify fruitlet and calyx structures.
ModelResolutionBounding BoxMask
  PRmAP 50mAP 50:95PRmAP 50mAP 50:95
YOLO11n-seg6400.6770.8510.8140.5120.6740.8480.8100.485
YOLO11n-seg9600.6860.8410.8350.5450.6850.8410.8340.523
YOLO11s-seg6400.6700.8290.7940.5190.6680.8260.7980.491
YOLO11s-seg9600.7530.7800.8170.5620.7530.7800.8200.540
YOLO11m-seg6400.6290.8900.8190.5380.6670.8840.8230.496
Table 16: Class-wise detection and instance-segmentation performance for the fruitlet class using YOLO11 models under 640 ×\times 640 and 960 ×\times 960 input resolutions.
ModelResolutionBounding BoxMask
  PRmAP 50mAP 50:95PRmAP 50mAP 50:95
YOLO11n-seg6400.6300.8150.7460.3720.6170.7990.7310.315
YOLO11n-seg9600.6680.8310.7590.3930.6700.8360.7710.370
YOLO11s-seg6400.6400.8150.7500.3820.6300.8020.7360.348
YOLO11s-seg9600.7130.7200.7440.4000.7160.7230.7430.369
YOLO11m-seg6400.5480.8600.7430.3930.5700.8430.7290.344
Table 17: Class-wise detection and instance-segmentation performance for the calyx class using YOLO11 models under 640 ×\times 640 and 960 ×\times 960 input resolutions.
ModelResolutionBounding BoxMask
  PRmAP 50mAP 50:95PRmAP 50mAP 50:95
YOLO11n-seg6400.3320.4040.3510.1880.3320.4040.3240.152
YOLO11n-seg9600.4320.4420.4040.2500.4320.4420.4010.207
YOLO11s-seg6400.3490.3610.3240.1900.3670.3800.3280.164
YOLO11s-seg9600.5400.4230.5210.3180.5400.4230.5120.297
YOLO11m-seg6400.6570.3460.4790.2800.7470.2840.4810.244
Table 18: Class-wise detection and instance-segmentation performance for the peduncle class using YOLO11 models under 640 ×\times 640 and 960 ×\times 960 input resolutions.
ClassΔ\Delta Box PΔ\Delta Box RΔ\Delta Box mAP 50Δ\Delta Box mAP 50:95Δ\Delta Mask PΔ\Delta Mask RΔ\Delta Mask mAP 50Δ\Delta Mask mAP 50:95
Fruitlet-0.005+0.011+0.014+0.033-0.011+0.015+0.015+0.042
Calyx+0.020+0.003+0.003+0.021+0.022+0.014+0.015+0.036
Peduncle+0.028+0.035+0.048+0.048+0.007+0.043+0.052+0.058
Table 19: Mean absolute change in class-wise YOLO11 performance when increasing input resolution from 640 ×\times 640 to 960 ×\times 960 pixels across all five model scales.
Figure 17: Qualitative comparison of YOLOv8 instance segmentation at 640×640640\times 640 and 960×960960\times 960 training resolutions. Solid red circles highlight peduncles missed by the 640×640640\times 640 model but successfully detected and segmented at 960×960960\times 960 , demonstrating improved preservation and recognition of fine-scale anatomical structures under higher-resolution training.
ModelResolutionBounding BoxMask
  PRmAP 50mAP 50:95PRmAP 50mAP 50:95
YOLOv8n-seg6400.7340.7870.7750.4920.7260.7770.7700.458
YOLOv8n-seg9600.7650.7650.8230.5570.7710.7710.8240.532
YOLOv8s-seg6400.7490.7710.7920.5160.7470.7650.7820.477
YOLOv8s-seg9600.6840.8380.8040.5430.6890.8450.8070.523
YOLOv8m-seg6400.7240.7990.7590.4940.7240.8020.7590.476
Table 20: Class-wise detection and instance-segmentation performance for the fruitlet class using YOLOv8 models under 640 ×\times 640 and 960 ×\times 960 input resolutions.
ModelResolutionBounding BoxMask
  PRmAP 50mAP 50:95PRmAP 50mAP 50:95
YOLOv8n-seg6400.6910.7480.7120.3280.6910.7480.6910.300
YOLOv8n-seg9600.6910.7550.7060.3620.6960.7610.7140.337
YOLOv8s-seg6400.7220.7640.7210.3580.7150.7550.7030.332
YOLOv8s-seg9600.6770.8310.7200.3760.6710.8250.7080.351
YOLOv8m-seg6400.6870.7710.6580.3570.6960.7830.6610.315
Table 21: Class-wise detection and instance-segmentation performance for the calyx class using YOLOv8 models under 640 ×\times 640 and 960 ×\times 960 input resolutions.
ModelResolutionBounding BoxMask
  PRmAP 50mAP 50:95PRmAP 50mAP 50:95
YOLOv8n-seg6400.4590.6540.4930.2490.4730.6730.4840.216
YOLOv8n-seg9600.4020.3650.4090.2510.4020.3650.4190.243
YOLOv8s-seg6400.4970.4420.4850.2950.4990.4420.4580.259
YOLOv8s-seg9600.4550.6540.5510.3380.4550.6540.5560.296
YOLOv8m-seg6400.4730.6350.4600.2760.4450.6000.4070.220
Table 22: Class-wise detection and instance-segmentation performance for the peduncle class using YOLOv8 models under 640 ×\times 640 and 960 ×\times 960 input resolutions.
ClassΔ\Delta Box PΔ\Delta Box RΔ\Delta Box mAP 50Δ\Delta Box mAP 50:95Δ\Delta Mask PΔ\Delta Mask RΔ\Delta Mask mAP 50Δ\Delta Mask mAP 50:95
Fruitlet+0.011-0.001+0.026+0.035+0.015-0.001+0.029+0.050
Calyx+0.023-0.006+0.013+0.023+0.031-0.002+0.018+0.030
Peduncle+0.021-0.044+0.023+0.035+0.031-0.039+0.051+0.045
Table 23: Mean absolute change in class-wise YOLOv8 performance when increasing input resolution from 640 ×\times 640 to 960 ×\times 960 pixels across all five model scales.
ModelΔ\Delta PΔ\Delta RΔ\Delta F1Δ\Delta Box mAP 50Δ\Delta Box mAP 50:95Δ\Delta Mask mAP 50Δ\Delta Mask mAP 50:95
YOLOv8
YOLOv8n-seg-0.009-0.102-0.052-0.014+0.034+0.003+0.046
YOLOv8s-seg-0.051+0.115+0.022+0.026+0.029+0.042+0.034
YOLOv8m-seg+0.012-0.099-0.039+0.034+0.019+0.045+0.034
YOLOv8l-seg+0.100-0.061+0.043+0.033+0.034+0.039+0.050
YOLOv8x-seg+0.039+0.059+0.048+0.023+0.039+0.032+0.045
Table 24: Paired change in aggregate detection and instance-segmentation performance when increasing input resolution from 640 ×\times 640 to 960 ×\times 960 pixels. Changes are calculated as ΔM=M960−M640\Delta M=M_{960}-M_{640} . F1 was derived from aggregate bounding-box precision and recall. Positive values denote improvement under the 960-pixel configuration.
ScaleΔ\Delta PΔ\Delta RΔ\Delta F1Δ\Delta Box mAP 50Δ\Delta Box mAP 50:95Δ\Delta Mask mAP 50Δ\Delta Mask mAP 50:95
Nano (n)+0.044-0.031+0.011+0.012+0.041+0.027+0.048
Small (s)+0.039+0.064+0.049+0.052+0.046+0.064+0.053
Medium (m)-0.003-0.023-0.013+0.009+0.015+0.021+0.030
Large (l)-0.005-0.004-0.002+0.004+0.022+0.020+0.037
Extra-large (x)+0.032+0.000+0.015+0.011+0.027+0.013+0.030
Table 25: Mean resolution-induced performance change for each model scale across YOLOv8, YOLO11, and YOLO26. Values represent the mean ΔM\Delta M for the three architecture generations.
GenerationΔ\Delta PΔ\Delta RΔ\Delta F1Δ\Delta Box mAP 50Δ\Delta Box mAP 50:95Δ\Delta Mask mAP 50Δ\Delta Mask mAP 50:95
YOLOv8+0.018-0.018+0.004+0.020+0.031+0.032+0.042
YOLO11+0.014+0.017+0.012+0.022+0.034+0.027+0.045
YOLO26+0.032+0.004+0.020+0.011+0.026+0.028+0.031
Table 26: Mean resolution-induced performance change for each YOLO generation across the five n/s/m/l/x model scales.
ModelTrain 640 (h)Train 960 (h)Δ\Delta Train (h)Train change (%)Δ\Delta Infer. (ms)Infer. change (%)
YOLOv8
YOLOv8n-seg0.4620.792+0.330+71.4+0.5+41.7
YOLOv8s-seg0.3690.724+0.355+96.2-0.4-14.3
YOLOv8m-seg0.4060.615+0.209+51.5+0.8+14.8
YOLOv8l-seg0.3320.791+0.459+138.3+4.7+104.4
YOLOv8x-seg0.5001.214+0.714+142.8+5.7+96.6
Table 27: Paired changes in training duration and validation-time inference latency between 640 ×\times 640 and 960 ×\times 960 configurations. Percentage changes are relative to the corresponding 640-pixel configuration. Positive values indicate increased computational cost.
Figure 18: Resolution-induced change in aggregate detection and segmentation accuracy across 15 YOLO model configurations. Each heatmap cell represents Δ\Delta mAP = mAP 960 –mAP 640 for the corresponding model and performance metric. Positive values indicate improved performance under the 960 ×\times 960 small-object-focused configuration, whereas negative values indicate reduced performance. The strongest and most consistent gains occur for the small-scale YOLO11s and YOLO26s configurations, while several large and extra-large configurations exhibit diminishing returns.
Figure 19: Resolution-induced change in mask mAP 50:95 across YOLO generations and model scales. Each point represents ΔmAP50:95=mAP50:95960−mAP50:95640\Delta\mathrm{mAP}_{50:95}=\mathrm{mAP}_{50:95}^{960}-\mathrm{mAP}_{50:95}^{640} . Positive values indicate improved instance-segmentation performance when increasing input resolution from 640 ×\times 640 to 960 ×\times 960 pixels, whereas negative values indicate reduced performance. YOLO11s exhibits the largest improvement ( Δ=+0.068\Delta=+0.068 ), while YOLO26x shows a slight reduction ( Δ=−0.003\Delta=-0.003 ), demonstrating that the effect of increased spatial resolution depends jointly on architecture generation and model scale.
GenerationModelLayersParameters (M)GFLOPsModel size (MB)
YOLOv8YOLOv8n-seg863.25911.36.8
YOLOv8s-seg8611.78139.923.9
YOLOv8m-seg10627.224104.354.8
YOLOv8l-seg12645.914210.192.3
YOLOv8x-seg12671.724328.0144.0
YOLO11YOLO11n-seg1142.8359.66.0
Table 28: Cross-generation architectural complexity and stored model size of the 15 YOLO instance-segmentation architectures. Parameters and GFLOPs correspond to the fused Ultralytics inference models. Model size corresponds to the optimizer-stripped best-performing checkpoint. The architecture is identical for the corresponding 640- and 960-pixel configurations.
  640 ×\times 640960 ×\times 960
Gen.ModelPreInfer.PostTotalPreInfer.PostTotal
YOLO26n0.11.90.32.30.21.80.62.6
s0.31.80.42.50.23.00.84.0
m0.13.00.43.50.25.90.86.9
l0.23.90.34.40.27.70.68.5
x0.16.30.97.30.212.80.713.7
Table 29: Per-image validation processing time for YOLOv8, YOLO11, and YOLO26 under 640 ×\times 640 and 960 ×\times 960 configurations. Total time is calculated as preprocessing + inference + postprocessing. All values are milliseconds per image.
GenerationModel640 ×\times 640960 ×\times 960
  CompletedBestTime (h)CompletedBestTime (h)
YOLO26n200–0.933200–1.073
s200–1.019200–1.076
m200–0.998200–1.257
l200–1.051200–1.409
x200–1.034200–2.124
Table 30: Training duration and convergence behavior of all configurations. “–” indicates that the full 200-epoch schedule was completed and no early-stopping best epoch was reported in the corresponding summary. YOLO26 completed all 200 epochs in every configuration.

Explore similar work

CardsList

  1. Comparing YOLOv11 and YOLOv8 for instance segmentation of occluded and non-occluded immature green fruits in complex orchard environment

    Oct 24, 2024Ranjan Sapkota, Manoj KarkeeAgricultural Image AnalysisInstance Segmentation

    Washington State University Department of Biological Systems Engineering

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

    Apr 11, 2026Mohammed Asad, Ajai Kumar Gautam, Priyanshu Dhiman +1Object DetectionAgricultural Image Analysis

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

  3. 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 AnalysisFine-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