ROSE-X: an annotated data set for evaluation of 3D plant organ segmentation methods.

Area of Science:

• Agricultural Science
• Computer Vision
• Plant Biology

Background:

• Annotated 3D plant datasets are crucial for developing and validating automatic phenotyping tools.
• Complex plant structures, like rosebushes, present significant annotation challenges for 3D vision-based phenotyping.

Purpose of the Study:

• Introduce the ROSE-X dataset, comprising 11 annotated 3D rosebush models.
• Provide ground truth data for training and benchmarking organ-level segmentation algorithms.
• Facilitate advancements in 3D plant phenotyping.

Main Methods:

• Acquired 3D models of real rosebush plants using X-ray tomography.
• Manually annotated voxels for organ-level labeling (ground truth).
• Presented data in both volumetric and point cloud formats.

Main Results:

• The ROSE-X dataset includes 11 high-quality, complex 3D rosebush models.
• Baseline leaf and stem segmentation achieved Intersection of Union (IoU) of 97.93% and 86.23%, respectively, using a volumetric approach with random forest classification.
• Identified challenges in segmenting touching organs in complex plant architectures.

Conclusions:

• The ROSE-X dataset provides a valuable resource for training and evaluating plant organ segmentation methods.
• Baseline segmentation results highlight areas for future methodological improvements.
• This dataset is poised to become a significant resource for automatic plant phenotyping research.