DBGCN: Dual-branch Graph Convolutional Network for organ instance inference on sparsely labeled 3D plant data.

Area of Science:

• Computer Vision
• Plant Science
• Machine Learning

Background:

• 3D crop phenotyping is crucial for plant gene screening and germplasm identification.
• Organ segmentation in 3D crop point clouds is a key but labor-intensive step.
• Current inductive deep learning methods require extensive manual data annotation, hindering progress.

Purpose of the Study:

• To develop a Graph Neural Network (GNN) based method for efficient organ segmentation in sparsely annotated 3D crop point clouds.
• To overcome the limitations of manual data labeling in 3D crop phenotyping.

Main Methods:

• Proposed a Dual-branch Graph Convolutional Network (DBGCN) utilizing transductive learning.
• DBGCN incorporates both static and dynamic graph convolutional operations on point clouds.
• The network performs organ instance inference directly on featureless point clouds with sparse labels.

Main Results:

• DBGCN achieved high node classification accuracy, outperforming mainstream GNNs and inductive deep architectures.
• Achieved 93.00% mAcc on PlantNet with 1.95% manual annotation and 91.05% mAcc on Soybean-MVS with 4.88% manual annotation.
• Demonstrated effectiveness on crop 3D data and potential for other point cloud segmentation tasks like street view.

Conclusions:

• DBGCN offers an efficient solution for 3D crop organ segmentation with significantly reduced annotation requirements.
• The fusion of static and dynamic graph convolutions enhances segmentation performance.
• The developed method shows broad applicability beyond crop phenotyping.