[Paper Review] Improved YOLOv12 with LLM-Generated Synthetic Data for Enhanced Apple Detection and Benchmarking Against YOLOv11 and YOLOv10

Ranjan Sapkota, Manoj Karkee|ArXiv.org|Feb 26, 2025

Plant and Fungal Interactions Research4 citations

TL;DR

YOLOv12 trained on synthetic data generated by LLMs outperforms YOLOv11 and YOLOv10 in apple detection, achieving higher precision, recall, and mAP@50, with field testing validating practicality.

ABSTRACT

This study evaluated the performance of the YOLOv12 object detection model, and compared against the performances YOLOv11 and YOLOv10 for apple detection in commercial orchards based on the model training completed entirely on synthetic images generated by Large Language Models (LLMs). The YOLOv12n configuration achieved the highest precision at 0.916, the highest recall at 0.969, and the highest mean Average Precision (mAP@50) at 0.978. In comparison, the YOLOv11 series was led by YOLO11x, which achieved the highest precision at 0.857, recall at 0.85, and mAP@50 at 0.91. For the YOLOv10 series, YOLOv10b and YOLOv10l both achieved the highest precision at 0.85, with YOLOv10n achieving the highest recall at 0.8 and mAP@50 at 0.89. These findings demonstrated that YOLOv12, when trained on realistic LLM-generated datasets surpassed its predecessors in key performance metrics. The technique also offered a cost-effective solution by reducing the need for extensive manual data collection in the agricultural field. In addition, this study compared the computational efficiency of all versions of YOLOv12, v11 and v10, where YOLOv11n reported the lowest inference time at 4.7 ms, compared to YOLOv12n's 5.6 ms and YOLOv10n's 5.9 ms. Although YOLOv12 is new and more accurate than YOLOv11, and YOLOv10, YOLO11n still stays the fastest YOLO model among YOLOv10, YOLOv11 and YOLOv12 series of models. (Index: YOLOv12, YOLOv11, YOLOv10, YOLOv13, YOLOv14, YOLOv15, YOLOE, YOLO Object detection)

Motivation & Objective

• Motivate robust apple detection in complex orchard environments while reducing reliance on costly field data collection.
• Assess the performance gains of YOLOv12 over YOLOv11 and YOLOv10 using synthetic data.
• Validate the model with real-field images to demonstrate practical applicability in agricultural automation.

Proposed method

• Generate synthetic apple orchard images using an LLM-based pipeline with DALL·E 2 and CLIP embeddings.
• Annotate synthetic images and train four YOLOv12 configurations (n, s, m, l) with fixed training hyperparameters.
• Compare YOLOv12 configurations against YOLOv11 and YOLOv10 using identical synthetic datasets and standard metrics (Precision, Recall, mAP@50).
• Evaluate computational efficiency including parameters, GFLOPs, and inference latency across models.
• Perform field testing by inference on real orchard images captured with a Kinect DK mounted on a robotic platform to assess generalization.

Experimental results

Research questions

• RQ1 Can YOLOv12 trained exclusively on LLM-generated synthetic data surpass YOLOv11 and YOLOv10 in apple detection metrics?
• RQ2 Which YOLOv12 configuration offers the best trade-off between accuracy and efficiency for orchard deployment?
• RQ3 Do synthetic-data-trained models generalize well to real-field orchard images?

Key findings

• YOLOv12n achieves the highest metrics among all configurations with Precision 0.916, Recall 0.969, and mAP@50 0.978.
• YOLOv12s/m/l achieve Precision 0.898, Recall 0.956, and mAP@50 0.974.
• YOLOv11x yields Precision 0.857, Recall 0.85, and mAP@50 0.91 as the top performer in the YOLOv11 series; YOLOv10n achieves Precision 0.84, Recall 0.8, and mAP@50 0.89 in the YOLOv10 series.
• YOLOv11n records the fastest inference at 4.7 ms, while YOLOv12n is 5.6 ms and YOLOv10n is 5.9 ms, indicating faster times for the older YOLOv11 variant.
• YOLOv12n uses the fewest parameters (2.556M) and 6.3 GFLOPs with 159 convolutional layers, suggesting a leaner, efficient architecture.
• Field testing with real images showed YOLOv12 surpasses predecessors in field conditions, confirming practical viability of LLm-generated synthetic data for agricultural deployment.