Host: The Japan Society of Mechanical Engineers
Name : [in Japanese]
Date : May 27, 2020 - May 30, 2020
This paper describes a method of dataset generation to realize cucumber recognition for harvesting robots by deep learning. To recognize cucumber fruits, conventional image processing methods are not very effective because cucumber fruits, leaves, and stems have same color, while deep learning is useful because it can automatically learn features to recognize them. Deep learning, however, requires a large amount of data for training, and there are no datasets for the cucumber greenhouse. Manual annotation consumes time and cost. Therefore, we developed a method to generate a large dataset automatically using computer graphics. We generated a dataset by creating and rendering 3D models of cucumber plants with their actual parameters measured in the greenhouse. Using this method, we made a cucumber dataset for semantic segmentation, object detection, and instance segmentation.
