Eff-3DPSeg: 3D Organ-Level Plant Shoot Segmentation Using Annotation-Efficient Deep Learning

Area of Science:

• Computer Vision
• Plant Science
• Machine Learning

Background:

• Accurate 3D plant shoot segmentation is crucial for extracting organ-level phenotypic traits.
• Fully supervised deep learning demands extensive, time-consuming point-wise annotations.
• Developing efficient annotation strategies is vital for advancing plant phenotyping.

Purpose of the Study:

• To propose a novel weakly supervised framework, Eff-3DPSeg, for automated 3D plant shoot segmentation.
• To reduce the annotation burden in deep learning models for plant organ segmentation.
• To enable accurate extraction of plant phenotypic traits for improved plant breeding.

Main Methods:

• Reconstructed high-resolution 3D point clouds of soybean using low-cost photogrammetry.
• Developed a Meshlab-based Plant Annotator for efficient point cloud annotation.
• Implemented a weakly supervised deep learning approach involving self-supervised pretraining and minimal fine-tuning (0.5% annotated points).

Main Results:

• Eff-3DPSeg achieved high precision (95.1%), recall (96.6%), F1 score (95.8%), and mIoU (92.2%) for soybean stem-leaf segmentation.
• Demonstrated comparable performance to fully supervised methods with significantly reduced annotation effort.
• Successfully extracted key phenotypic traits including stem diameter, leaf width, and leaf length.

Conclusions:

• Weakly supervised deep learning offers an effective solution for 3D plant shoot segmentation, minimizing annotation costs.
• The Eff-3DPSeg framework provides a scalable approach for characterizing 3D plant architecture.
• This method has the potential to accelerate plant breeding by enhancing selection processes through detailed phenotyping.