In recent years, deep learning has made significant advancements in various fields, and now it is being utilized in the realm of plant studies as well. The integration of deep learning techniques with point clouds has shown remarkable progress in the 3D segmentation of plant shoots. Unlike traditional 2D methods, which faced challenges in accurately perceiving depth and determining structures, 3D imaging has overcome these limitations and provided better analysis of plant phenotypic traits.
However, one of the challenges of 3D imaging is the need to carefully label each point in the image, which is a time-consuming and expensive task. To address this issue, researchers have been exploring the use of supervised learning models that require fewer labeled points.
A recent study named Eff-3DPSeg: 3D Organ-Level Plant Shoot Segmentation Using Annotation-Efficient Deep Learning introduces a weakly supervised deep learning framework for plant organ segmentation. The researchers developed this framework by utilizing a Multi-view Stereo Pheno Platform (MVSP2) to acquire point clouds from individual plants. These point clouds were then annotated using a Meshlab-based Plant Annotator (MPA).
The framework involves two main steps. First, the researchers reconstructed high-resolution point clouds of soybean plants using a cost-effective photogrammetry system. They also developed the Meshlab-based Plant Annotator to annotate the plant point clouds. Then, they employed a weakly supervised deep learning method for plant organ segmentation. The model was pretrained with only approximately 0.5 percent of labeled points and fine-tuned using the Viewpoint Bottleneck loss to learn intrinsic structure representation from raw point clouds. From the segmented plant organ, three phenotypic traits were extracted: the length and width of leaves, and the diameter of the stem.
To evaluate the performance of the framework, the researchers tested it on different growth stages using a large soybean spatiotemporal dataset and compared the results with fully labeled techniques on tomato and soybean plants. While the stem-leaf segmentation results were generally accurate, some misclassifications were observed at leaf edges and junctions. The approach also performed better on less complex plant structures and achieved greater accuracy with larger training sets. Notably, quantitative results demonstrated significant improvements over baseline techniques, particularly in less supervised environments.
Despite these advancements, the study encountered limitations such as data gaps and the need for separate training for different segmentation tasks. The researchers acknowledge these limitations and emphasize the need for future refinement of the framework, including expansion to a wider range of plant classifications and growth phases.
In conclusion, the Eff-3DPSeg framework represents a significant advancement in 3D plant shoot segmentation. Its efficient annotation process and accurate segmentation capabilities have the potential to enhance high throughput in plant studies. Moreover, by employing weakly supervised deep learning and innovative annotation techniques, Eff-3DPSeg overcomes the challenges of expensive and time-consuming labeling processes. This framework paves the way for future developments in the field of plant segmentation and analysis.