A Comparative Study of Physically Accurate Synthetic Shadow Datasets in Agricultural Settings with Human Activity

Area of Science:

• Computer Vision
• Agricultural Technology
• Machine Learning

Background:

• Shadows are crucial in agriculture, impacting plant photosynthesis.
• Existing shadow datasets often have annotation errors and lack agricultural context, especially with human activity.
• Top-down views are underrepresented in current shadow datasets.

Purpose of the Study:

• To evaluate a synthetically generated, photorealistic top-down shadow segmentation dataset for agriculture.
• To compare the synthetic dataset's efficacy against real-world datasets for neural network training.
• To assess the influence of annotation quality and image domain on model performance.

Main Methods:

• Generation of a synthetic top-down shadow segmentation dataset with accurate masks.
• Training baseline neural network architectures on the synthetic dataset.
• Exploration of transfer learning using various public shadow datasets.
• Evaluation of out-of-domain performance against other shadow datasets.

Main Results:

• The synthetic dataset, AgroSegNet, demonstrates competitive performance in shadow segmentation.
• AgroSegNet is effective for transfer learning, especially in agriculture-related domains.
• Findings highlight the potential of high-quality synthetic data for agricultural computer vision.

Conclusions:

• High-quality synthetic data can effectively train neural networks for agricultural shadow segmentation.
• AgroSegNet offers a valuable resource for advancing research in agricultural computer vision.
• The study underscores the importance of domain relevance and annotation quality in dataset creation.