Structure-Aware Completion of Plant 3D LiDAR Point Clouds via a Multi-Resolution GAN-Inversion Network

Zhiming Wei^1,2Jianing Long³Zhihong Zhang⁴Xinyu Xue^2*Yitian Sun¹Qinglong Li¹Wu Liu⁵Jingxin Shen¹Zhikai Zhang⁶Xiaoju Li⁶Zhengguo Ma⁶

• ¹Shandong Academy of Agricultural machinery Sciences, Jinan, China
• ²Nanjing Institute of Agricultural Mechanization Ministry of Agriculture and Rural Affairs, Nanjing, China
• ³China Agricultural University, Beijing, China
• ⁴Shanghai Institute of Technology, Shanghai, China
• ⁵Bureau of Agriculture andRural Affairs of Yanggu County, Liaocheng, China
• ⁶Qingdao Agricultural University, Qingdao, China

Three-dimensional (3D) point cloud data acquired by sensors such as Light Detection and Ranging (LiDAR) are foundational across diverse application domains, including precision agriculture, mobile robotics, autonomous driving, construction and infrastructure inspection, and cultural heritage documentation. However, environmental disturbances and sensor limitations frequently lead to incomplete or noisy data, which significantly impairs the accuracy of subsequent applications. To address this challenge, this research introduces a novel unsupervised deep learning method, the Multi-Resolution Completion Net (MRC-Net), designed to effectively complete imperfect point cloud data while preserving both global structure and intricate details. MRC-Net, an advancement on the ShapeInversion method, ingeniously integrates a Generative Adversarial Network (GAN) inversion approach with multi-resolution principles. Its core architecture features an encoder for feature extraction, a generator for data completion, and a discriminator to validate data integrity and detail. The model's key innovations are its multi-resolution degradation mechanism and a multi-scale discriminator, which functions synergistically to capture both macro-structures and micro-details. To comprehensively evaluate its performance, the model was tested extensively on various datasets. On virtual datasets such as CRN, MRC-Net's performance was comparable to leading supervised networks, achieving an average Chamfer Distance (CD) of 8.0 and an F1 score of 91.3. More significantly, on a custom dataset designed to simulate agricultural applications, MRC-Net demonstrated exceptional practical utility. When completing standard objects like regular cartons, it achieved a CD of 3.3 and an F1 score of 97.3. For structurally complex simulated plants, the model successfully maintained the integrity of the overall structure despite minor imperfections, yielding average CD and F1 scores of 8.6 and 88.1, respectively. The experimental results confirm that MRC-Net represents a significant advance in point cloud completion. This study holds the potential to provide a reliable and accurate data foundation for downstream applications such as autonomous navigation, high-precision 3D modeling, and agricultural robotics, thereby making a valuable contribution to enhancing data quality in precision agriculture.