Abstract:
Plant research has primarily focused on above-ground traits, such as leaves and flowers, while roots have received comparatively less attention due to their difficulty in imaging. Minirhizotron (MR) systems are commonly used to capture root images underground, but their use may impact plant root growth, and they only provide a two-dimensional (2D) view of the root architecture. X-ray computed tomography (CT), on the other hand, captures three-dimensional (3D) information about soil cores extracted from the soil. This work aims to refine the segmentation of pre-segmented results, which form a set of points (called point clouds), as input. However, current approaches to point cloud semantic segmentation for plant roots typically require manual annotation of the soil core point-by-point and utilize supervised graph convolutional networks (GCN). Such approaches have several limitations, including the difficulty of annotating 3D point clouds, a focus on indoor objects with high interclass variances and low intraclass variances, and a lack of consideration for varying point cloud densities at different depths in plant root point clouds. To address these limitations, this work proposes a weakly supervised point cloud semantic segmentation approach using GCNs that take point cloud densities into account. The proposed methods are evaluated on three challenging root datasets and compared to several baseline approaches.
Links:
Citation:
Y. Cui, “Weakly Supervised Point Cloud Semantic Segmentation With Graph Convolutional Networks.” University of Florida, 2023.@book{cui2023weakly,
title={Weakly Supervised Point Cloud Semantic Segmentation with Graph Convolutional Networks},
author={Cui, Yiming},
year={2023},
publisher={University of Florida}
}