Related Experiment Videos
Accurate orange yield estimation using a novel dataset, fine-tuned deep learning models, and vision-LLM benchmarking
Qurat Ul Ain Akram1, Zeeshan Ramzan1, Waqas Ali2
1Department of Computer Science, New Campus, University of Engineering and Technology, Lahore, Pakistan.
Scientific Reports
|July 10, 2026
Summary
This study introduces a new dataset and methods for accurate orange yield estimation, addressing challenges like occlusion and varying conditions. Faster R-CNN achieved the best results, improving precision agriculture for fruit farming.
Area of Science:
- Computer Vision and Machine Learning in Agriculture
- Precision Agriculture Technologies
- Fruit Yield Estimation using Deep Learning
Background:
- Traditional manual orange yield estimation is labor-intensive, time-consuming, and error-prone.
- Deep learning offers potential for enhanced fruit detection and yield estimation in precision agriculture.
- Challenges in orange yield estimation include fruit density, occlusion, lighting variations, and image quality.
Purpose of the Study:
- To develop a comprehensive dataset for orange yield estimation, accounting for variations in size, lighting, and occlusion.
- To benchmark state-of-the-art deep learning models for accurate orange detection and yield estimation.
- To evaluate the performance of Vision Large Language Models (VLLMs) in zero-shot orange yield estimation.
Main Methods:
- Collected a dataset of 9,781 images from 41 videos, annotated with bounding boxes for single oranges and bunches (2-4 oranges).
- Fine-tuned three multiclass object detection models (Faster R-CNN, Mask R-CNN, YOLOv8n) on the annotated dataset.
- Performed zero-shot evaluation of Vision Large Language Models (Gemini and LLaMA) for yield estimation.
Main Results:
- Faster R-CNN achieved the best performance with a Mean Absolute Percentage Error (MAPE) of 18.22% and an R-squared value of 0.95.
- Gemini and LLaMA showed promising zero-shot capabilities with MAPE values of 52.05% and 103.77%, respectively.
- The developed dataset and annotation strategy effectively handle occluded oranges and variations in image conditions.
Conclusions:
- The created dataset is valuable for advancing research in orange detection and yield estimation.
- Faster R-CNN demonstrates superior performance for orange yield estimation compared to Mask R-CNN and YOLOv8n.
- Further research is needed to improve the accuracy of VLLMs for practical agricultural applications.
