Standardizing and Centralizing Datasets for Efficient Training of Agricultural Deep Learning Models

Area of Science:

• Agricultural Computer Vision
• Deep Learning
• Machine Learning

Background:

• Deep learning models are standard in agricultural computer vision but often use general datasets for fine-tuning.
• This approach can lead to increased training time, resource use, and decreased model performance, reducing data efficiency.

Purpose of the Study:

• To enhance data efficiency in training agricultural deep learning models.
• To develop and evaluate methods for improving model performance and reducing training time without major pipeline changes.

Main Methods:

• Collected and standardized diverse public agricultural datasets for three distinct tasks.
• Established standard training and evaluation pipelines with benchmarks and pre-trained models.
• Experimented with novel deep learning methods and domain-specific agricultural applications.

Main Results:

• Agricultural pre-trained model weights and spatial data augmentations significantly improve model performance and reduce convergence time.
• Models trained on low-quality annotations achieve performance comparable to those trained on high-quality data.
• Methods are broadly applicable and show potential for substantial data efficiency gains.

Conclusions:

• Optimizing deep learning training strategies, even with minor modifications, can yield significant improvements in agricultural computer vision.
• The use of agricultural-specific pre-trained models and effective data augmentation are key to enhancing data efficiency.
• The findings suggest that lower-quality annotated datasets can still be valuable for training, expanding available data resources.