Sign in Subscribe
Agricultural industry

Minimum Dataset Requirements for Fine-Tuning Object Detectors in Arable Crop Plant Counting

Strategic Insights

— 1 min read

Investigations into precision agriculture applications have shown that object detection is crucial for accurate plant counting. However, researchers have struggled to identify the minimum dataset requirements for effective model deployment. A recent study published by the University of Turin shed light on this issue, specifically for maize seedling detection on orthomosaics. The research team evaluated traditional deep learning models, newer approaches, and methods requiring zero labeled examples, testing them with varying training sources, sizes, and annotation quality levels. The findings demonstrate that models trained on in-domain data achieved acceptable performance with as few as 60-130 annotated images, depending on architecture.

Key Takeaways:

  • The study systematically evaluated traditional deep learning models (YOLOv5, YOLOv8, YOLO11, RT-DETR), newer approaches (CD-ViTO), and methods requiring zero labeled examples (OWLv2) for maize seedling detection.
  • Models trained on in-domain data reached the benchmark with as few as 60-130 annotated images, depending on architecture.
  • Transformer-based models (RT-DETR) required significantly fewer samples (60) than CNN-based models (110-130), though they showed different tolerances to annotation quality reduction.
  • Models maintained acceptable performance with only 65-90% of original annotation quality.
  • Despite recent advances, neither few-shot nor zero-shot approaches met minimum performance requirements for precision agriculture deployment.
  • The research concluded that successful deployment requires in-domain training data, with minimum dataset requirements varying by model architecture.

Statistics:

  • 60-130 annotated images were required for models to achieve acceptable performance with in-domain data.
  • Transformer-based models (RT-DETR) required 60 samples, while CNN-based models required 110-130 samples.
  • 65-90% of original annotation quality was sufficient for most models to maintain acceptable performance.
  • The study did not find that even the most advanced models could meet minimum performance requirements with out-of-distribution data.

Sources:

  • On the Minimum Dataset Requirements for Fine-Tuning an Object Detector for Arable Crop Plant Counting: A Case Study on Maize Seedlings (Remote Sensing, 2025, 17(13):2190)
  • https://doi-org.sdpl.idm.oclc.org/10.3390/rs17132190
  • Department of Agricultural, Forest and Food Sciences, University of Turin, 10095 Grugliasco, Italy (Samuele Bumbaca, Enrico Borgogno-Mondino)
  • Remote Sensing is a publisher of MDPI AG.

Read more