In this paper, we consider semantic segmentation of 3D plant point clouds. We propose various modifications to improve the performance of two main state-of-the-art methods for semantic segmentation of plants. In the first method, we consider the classical machine learning approach where local geometric features are classified via SVM. We propose to process the point cloud in two stages: The points classified, in the first stage, with high confidence are removed to enable retraining of the SVM in the second stage. Furthermore, we apply a multi-resolution approach to feature extraction. The second state-of-the-art method is PointNet++, which is a widely used deep learning architecture for plant segmentation. We propose alternative approaches, based on critical points and guide points, for data preprocessing for PointNet++. We observed that these modifications improved the segmentation performance on the tomato models of the Pheno4D data set.