UnPlantPC: Unsupervised plant point cloud completion based on keypoint sampling and region-aware contrastive chamfer distance loss

作者：Xiaomeng Li, Fei Li, Yuan Ding, Haoxiang Sun, Zhenbo Li · 发表于：Computers and Electronics in Agriculture · 年份：2025 · DOI：10.1016/j.compag.2025.110547 · 被引用次数：2 · 研究领域：Remote Sensing and LiDAR Applications、3D Surveying and Cultural Heritage、3D Shape Modeling and Analysis

In smart agriculture, the precise acquisition of complete 3D plant phenotypic data is critical for applications such as intelligent breeding and growth monitoring. However, due to equipment constraints, environmental noise, and self-occlusion, the collected 3D point cloud data of plants is often incomplete. This incompleteness significantly hinders key tasks in plant phenotypic analysis, including organ segmentation and surface reconstruction, necessitating effective data completion methods. Supervised point cloud completion methods face challenges due to the inherent incompleteness of collected data and the need for extensive labeled datasets. To address these issues, we propose UnPlantPC, an unsupervised plant point cloud completion model built on a self-supervised encoder–decoder paradigm. To effectively capture regions with complex geometric structures in plant point clouds, the model employs a keypoint down-sampling strategy that integrates Euclidean and cosine distances, ensuring the extracted key points are both representative and directionally informative. Additionally, a geometric-aware attention module enhances feature extraction in these regions, further improving the model’s ability to capture intricate geometric details. To align plant point cloud distributions under self-supervised learning, we introduce a novel Region-Aware Contrastive Distance, which provides accurate supervisory information. This innovation enables the model to deliver more precise completion...