GCASSN: a graph convolutional attention synergistic segmentation network for 3D plant point cloud segmentation.

Area of Science:

• Agricultural Science
• Computer Vision
• Data Science

Background:

• 3D point clouds are crucial for plant phenotyping, overcoming limitations of 2D imaging like leaf occlusion.
• Accurate plant point cloud segmentation is essential for analyzing phenotypic traits.
• Existing methods struggle to balance lightweight design with high segmentation precision.

Purpose of the Study:

• To propose a novel network, the Graph Convolutional Attention Synergistic Segmentation Network (GCASSN), for efficient and precise plant point cloud segmentation.
• To integrate graph convolutional networks (GCNs) and self-attention mechanisms for comprehensive feature extraction.
• To establish a robust foundation for advanced plant phenotype analysis.

Main Methods:

• The GCASSN framework includes Trans-net for point cloud normalization and a Graph Convolutional Attention Synergistic Module (GCASM).
• GCASM combines GCNs for local feature extraction and self-attention for global context.
• The network is designed to balance computational efficiency with segmentation accuracy.

Main Results:

• GCASSN achieved state-of-the-art results on Plant3D and Phone4D datasets, with 95.46% mean accuracy and 90.41% mIoU.
• The model demonstrated strong generalizability on the ShapeNet dataset, achieving 85.47% mIoU without parameter tuning.
• Computational efficiency was competitive, with inference time and parameter count comparable to existing methods like DGCNN.

Conclusions:

• The GCASSN effectively extracts both local and global features from plant point clouds, enabling robust segmentation.
• This method provides a solid foundation for detailed plant phenotype analysis in agricultural research.
• The developed GCASSN offers a significant advancement in processing 3D plant data for phenotyping.