The Challenge of Data Annotation in Deep Learning-A Case Study on Whole Plant Corn Silage

Area of Science:

• Computer Vision
• Machine Learning
• Agricultural Science

Background:

• Deep learning for computer vision requires large datasets, which are resource-intensive to create.
• Existing large benchmark datasets are difficult to implement in smaller projects with limited resources.
• Annotator inconsistency can hinder model evaluation in computer vision tasks.

Purpose of the Study:

• To present a process for creating an image dataset for kernel fragmentation and stover overlengths in Whole Plant Corn Silage.
• To address challenges of occlusion and clutter in agricultural image datasets.
• To evaluate Semi-Supervised Learning (SSL) as an efficient alternative to manual annotation.

Main Methods:

• Developed annotation guidelines for object instances in Whole Plant Corn Silage images.
• Collected and analyzed statistics of gathered annotations.
• Evaluated models using physically based sieving metrics independent of manual annotation.
• Assessed the performance of Semi-Supervised Learning with varying amounts of annotated data.

Main Results:

• The created dataset is suitable for training models despite challenging image conditions.
• Annotator inconsistency was identified as a significant challenge.
• Semi-Supervised Learning demonstrated competitive results, significantly improving Average Precision.
• SSL achieved over 3x improvement in Average Precision with a small dataset (100 images).

Conclusions:

• Physically based evaluation metrics are crucial for robust model assessment, independent of manual annotation.
• Semi-Supervised Learning offers a viable and efficient alternative to traditional manual annotation for agricultural computer vision tasks.
• The developed dataset and SSL approach show promise for improving kernel fragmentation and stover overlength analysis in Whole Plant Corn Silage.