Structure-aware completion of plant 3D LiDAR point clouds via a multi-resolution GAN-inversion network.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• 3D point clouds from LiDAR are crucial for robotics and infrastructure inspection.
• Incomplete or noisy point clouds hinder the performance of downstream applications.
• Robust and high-fidelity point cloud completion is essential for practical use.

Purpose of the Study:

• To develop an unsupervised deep learning framework for accurate 3D point cloud completion.
• To address the challenge of noisy and incomplete data in LiDAR scans.
• To improve the reliability of 3D data for applications like autonomous navigation and precision agriculture.

Main Methods:

• Proposed Multi-Resolution Completion Net (MRC-Net), an unsupervised deep learning framework.
• Integrated Generative Adversarial Network (GAN) inversion with multi-resolution principles.
• Employed a multi-resolution degradation mechanism and a multi-scale discriminator for effective reconstruction.

Main Results:

• MRC-Net achieved accuracy comparable to supervised methods on virtual datasets (e.g., CD 8.0, F1 91.3).
• Demonstrated high performance on agricultural datasets, preserving object integrity (e.g., CD 3.3, F1 97.3 for cartons).
• Successfully maintained overall shape for complex objects like simulated plants (CD 8.6, F1 88.1).

Conclusions:

• MRC-Net effectively balances global shape consistency with fine-grained detail in point cloud completion.
• The method provides a reliable data foundation for critical downstream tasks.
• Advances unsupervised point cloud completion, benefiting precision agriculture and robotics.