Using Key Point Detection to Extract Three Dimensional Phenotypes of Corn

Corn 3D phenotype extraction faces several problems, including low precision, excessive manual involvement, long processing times, and the requirement for complex equipment. To address these issues, we propose a novel key point detection deep learning model called V7POSE-GSConV, which operates on RGB-D data. This method is built upon the YOLOv7-POSE key point detection model, allowing us to directly capture key points of corn plants using RGB data as input. Leveraging corresponding RGB-D data, this work can derive the comprehensive structure of corn, extract parameters such as leaf length, leaf angle, plant height, and ear phenotype. In our experiments, the YOLOv7-POSE model achieved a recognition accuracy of 99.45% in training stage. To further optimize the model for efficiency while maintaining accuracy, we introduced GSConV. The results demonstrate a 2% reduction in the number of model parameters in V7POSE-GSConV, with no loss in accuracy.