Click on “Download PDF” for the PDF version or on the title for the HTML version.

If you are not an ASABE member or if your employer has not arranged for access to the full-text, Click here for options.

Abstract. Three-dimensional (3D) high throughput plant phenotyping techniques provide an opportunity to measure and predict plant geometric traits and their responses to changing environmental factors, which is highly useful to speed up the selection of new genotypes under specific growing conditions. 3D plant shoot segmentation is essential to obtain plant phenotypic traits at the organ level. Convolutional neural networks have been used for plant point cloud segmentation. However, the network training needs point-wise plant annotation which is extremely expensive and time-consuming. In this work, we aimed to address the challenge of network training using a small subset of annotations, about only 0.5% of labeled points. To this end, we proposed a framework, Eff-PlantNet, which contains two stages. In the first stage, meaningful representations from the plant point clouds were learned through a 3D self-supervised representation learning network without the usage of annotation. In the second stage, a subsequent weakly-supervised fine-tuning by the pre-trained model was used to conduct point cloud segmentation. Our framework was low-cost and effective and produced similar plant semantic and instance segmentation performance compared with the full-supervised training network.