St. John's Wort Detection

Project overview

Agricultural field images were annotated for a computer vision dataset focused on detecting St. John's Wort in real-world field environments. Visible plants were identified and enclosed in accurate annotations to generate high-quality training data for an AI-based object detection model.

The main objective was to prepare a reliable dataset for automatically detecting St. John's Wort plants in agricultural and natural environments.

Work performed

• Reviewed and organised the agricultural field image dataset before annotation.
• Identified visible St. John's Wort plants in each image.
• Drew accurate bounding boxes around individual target plants.
• Applied the STW class consistently across the dataset.
• Annotated partially visible plants where applicable.
• Distinguished St. John's Wort carefully from surrounding crops, weeds, grass and other vegetation.
• Handled plants at a range of growth stages and sizes.
• Performed quality checks and corrected inaccurate, missing or poorly positioned annotations.
• Ensured the final annotations were consistent and suitable for object detection model training.

What made it hard

• Differentiating St. John's Wort from other weeds and surrounding vegetation.
• Small plants were hard to identify in dense field environments.
• Partially visible and overlapping plants complicated placement.
• Growth stages and plant sizes varied across the dataset.
• Complex backgrounds contained grass, crops, soil and other weeds.
• Varying lighting conditions and shadows affected plant visibility.
• Keeping bounding boxes tight and accurate around irregularly shaped plants required care.

Result

A high-quality St. John's Wort detection dataset was completed with accurate and consistent annotations under the STW class. The dataset was prepared for training a computer vision model capable of automatically detecting St. John's Wort in real-world agricultural and natural environments.

Final deliverable

Accurate and consistent bounding box annotations for the STW class were delivered across the dataset. The final output was prepared to the project requirements and was ready for AI model training and automated weed detection applications.