3D data-augmentation methods for semantic segmentation of tomato plant parts

Area of Science:

• Computer Vision
• Plant Science
• Machine Learning

Background:

• 3D semantic segmentation of plant point clouds is crucial for automated plant phenotyping and crop modeling.
• Deep learning methods require large annotated datasets, which are time-consuming and labor-intensive to create for 3D plant structures.
• Effective data augmentation strategies are needed to improve model performance with limited training data.

Purpose of the Study:

• To investigate and compare the effectiveness of novel and existing data augmentation methods for 3D semantic segmentation of plant point clouds.
• To identify which data augmentation techniques are most suitable for segmenting 3D tomato plant structures.

Main Methods:

• Proposed five novel data augmentation methods: global cropping, brightness adjustment, leaf translation, leaf rotation, and leaf crossover.
• Compared these novel methods against five existing techniques: online down sampling, global jittering, global scaling, global rotation, and global translation.
• Applied the augmentation methods to the PointNet++ model for segmenting 3D point clouds of three tomato cultivars into soil base, stick, stemwork, and other bio-structures.

Main Results:

• Leaf crossover emerged as the most promising novel method, outperforming existing techniques.
• Leaf rotation, leaf translation, and cropping also showed strong performance, surpassing most existing methods except for global jittering.
• The proposed 3D data augmentation approaches effectively mitigated overfitting caused by limited training data.

Conclusions:

• Novel data augmentation methods, particularly leaf crossover, significantly enhance 3D semantic segmentation accuracy for plant point clouds.
• These improved segmentation capabilities lead to more accurate plant architecture reconstruction, advancing automated plant phenotyping.
• The study highlights the importance of tailored data augmentation for deep learning in plant science applications.